☆ Diffusion Buffer for Online Generative Speech Enhancement

Bunlong Lay, Rostislav Makarov, Simon Welker, Maris Hillemann, Timo Gerkmann

Online Speech Enhancement was mainly reserved for predictive models. A key advantage of these models is that for an incoming signal frame from a stream of data, the model is called only once for enhancement. In contrast, generative Speech Enhancement models often require multiple calls, resulting in a computational complexity that is too high for many online speech enhancement applications. This work presents the Diffusion Buffer, a generative diffusion-based Speech Enhancement model which only requires one neural network call per incoming signal frame from a stream of data and performs enhancement in an online fashion on a consumer-grade GPU. The key idea of the Diffusion Buffer is to align physical time with Diffusion time-steps. The approach progressively denoises frames through physical time, where past frames have more noise removed. Consequently, an enhanced frame is output to the listener with a delay defined by the Diffusion Buffer, and the output frame has a corresponding look-ahead. In this work, we extend upon our previous work by carefully designing a 2D convolutional UNet architecture that specifically aligns with the Diffusion Buffer's look-ahead. We observe that the proposed UNet improves performance, particularly when the algorithmic latency is low. Moreover, we show that using a Data Prediction loss instead of Denoising Score Matching loss enables flexible control over the trade-off between algorithmic latency and quality during inference. The extended Diffusion Buffer equipped with a novel NN and loss function drastically reduces the algorithmic latency from 320 - 960 ms to 32 - 176 ms with an even increased performance. While it has been shown before that offline generative diffusion models outperform predictive approaches in unseen noisy speech data, we confirm that the online Diffusion Buffer also outperforms its predictive counterpart on unseen noisy speech data.

☆ Adapting Language Balance in Code-Switching Speech ICASSP 2026

Enes Yavuz Ugan, Ngoc-Quan Pham, Alexander Waibel

Despite achieving impressive results on standard benchmarks, large foundational models still struggle against code-switching test cases. When data scarcity cannot be used as the usual justification for poor performance, the reason may lie in the infrequent occurrence of code-switched moments, where the embedding of the second language appears subtly. Instead of expecting the models to learn this infrequency on their own, it might be beneficial to provide the training process with labels. Evaluating model performance on code-switching data requires careful localization of code-switching points where recognition errors are most consequential, so that the analysis emphasizes mistakes occurring at those moments. Building on this observation, we leverage the difference between the embedded and the main language to highlight those code-switching points and thereby emphasize learning at those locations. This simple yet effective differentiable surrogate mitigates context bias during generation -- the central challenge in code-switching -- thereby improving the model's robustness. Our experiments with Arabic and Chinese-English showed that the models are able to predict the switching places more correctly, reflected by the reduced substitution error.

comment: Submitted to ICASSP 2026

☆ Bayesian Low-Rank Factorization for Robust Model Adaptation ICASSP 2026

Enes Yavuz Ugan, Ngoc-Quan Pham, Alexander Waibel

Large speech foundation models achieve strong performance across many domains, but they often require adaptation to handle local needs such as code-switching, where speakers mix languages within the same utterance. Direct fine-tuning of these models risks overfitting to the target domain and overwriting the broad capabilities of the base model. To address this challenge, we explore Bayesian factorized adapters for speech foundation models, which place priors near zero to achieve sparser adaptation matrices and thereby retain general performance while adapting to specific domains. We apply our approach to the Whisper model and evaluate on different multilingual code-switching scenarios. Our results show only minimal adaptation loss while significantly reducing catastrophic forgetting of the base model. Compared to LoRA, our method achieves a backward gain of 54% with only a 4% drop on the new domain. These findings highlight the effectiveness of Bayesian adaptation for fine-tuning speech foundation models without sacrificing generalization.

comment: Submitted to ICASSP 2026

☆ MLMA: Towards Multilingual with Mamba Based Architectures ICASSP 2026

Mohamed Nabih Ali, Daniele Falavigna, Alessio Brutti

Multilingual automatic speech recognition (ASR) remains a challenging task, especially when balancing performance across high- and low-resource languages. Recent advances in sequence modeling suggest that architectures beyond Transformers may offer better scalability and efficiency. In this work, we introduce MLMA (Multilingual Language Modeling with Mamba for ASR), a new approach that leverages the Mamba architecture--an efficient state-space model optimized for long-context sequence processing--for multilingual ASR. Using Mamba, MLMA implicitly incorporates language-aware conditioning and shared representations to support robust recognition across diverse languages. Experiments on standard multilingual benchmarks show that MLMA achieves competitive performance compared to Transformer-based architectures. These results highlight Mamba's potential as a strong backbone for scalable, efficient, and accurate multilingual speech recognition.

comment: The paper is under review at ICASSP 2026

☆ Noise-Conditioned Mixture-of-Experts Framework for Robust Speaker Verification

Bin Gu, Lipeng Dai, Huipeng Du, Haitao Zhao, Jibo Wei

Robust speaker verification under noisy conditions remains an open challenge. Conventional deep learning methods learn a robust unified speaker representation space against diverse background noise and achieve significant improvement. In contrast, this paper presents a noise-conditioned mixture-ofexperts framework that decomposes the feature space into specialized noise-aware subspaces for speaker verification. Specifically, we propose a noise-conditioned expert routing mechanism, a universal model based expert specialization strategy, and an SNR-decaying curriculum learning protocol, collectively improving model robustness and generalization under diverse noise conditions. The proposed method can automatically route inputs to expert networks based on noise information derived from the inputs, where each expert targets distinct noise characteristics while preserving speaker identity information. Comprehensive experiments demonstrate consistent superiority over baselines, confirming that explicit noise-dependent feature modeling significantly enhances robustness without sacrificing verification accuracy.

☆ A Stage-Wise Learning Strategy with Fixed Anchors for Robust Speaker Verification

Bin Gu, Lipeng Dai, Huipeng Du, Haitao Zhao, Jibo Wei

Learning robust speaker representations under noisy conditions presents significant challenges, which requires careful handling of both discriminative and noise-invariant properties. In this work, we proposed an anchor-based stage-wise learning strategy for robust speaker representation learning. Specifically, our approach begins by training a base model to establish discriminative speaker boundaries, and then extract anchor embeddings from this model as stable references. Finally, a copy of the base model is fine-tuned on noisy inputs, regularized by enforcing proximity to their corresponding fixed anchor embeddings to preserve speaker identity under distortion. Experimental results suggest that this strategy offers advantages over conventional joint optimization, particularly in maintaining discrimination while improving noise robustness. The proposed method demonstrates consistent improvements across various noise conditions, potentially due to its ability to handle boundary stabilization and variation suppression separately.

☆ ProLAP: Probabilistic Language-Audio Pre-Training

Toranosuke Manabe, Yuchi Ishikawa, Hokuto Munakata, Tatsuya Komatsu

Language-audio joint representation learning frameworks typically depend on deterministic embeddings, assuming a one-to-one correspondence between audio and text. In real-world settings, however, the language-audio relationship is inherently many-to-many: one audio segment can be described by multiple captions and vice versa. To address this, we propose Probabilistic Language-Audio Pre-training (ProLAP), which models multiplicity as the spread of probability distributions in a joint language-audio embedding space. To train the intra-modal hierarchical relationship effectively, we also introduce two objectives: (i) hierarchical inclusion loss to promote semantic hierarchical understanding of inputs and (ii) mask repulsive loss to improve the efficiency of learning when optimizing the hierarchical inclusion loss. With this training strategy, our model can learn the hierarchical structure inherent in the data even from small datasets, in contrast to prior probabilistic approaches that rely on large-scale datasets. In our experiments, ProLAP outperforms existing deterministic approaches on audio-text retrieval tasks. Moreover, through experiments on the audio traversal task introduced in this paper, we demonstrate that ProLAP captures the plausible semantic hierarchy.

comment: Under review

☆ SegTune: Structured and Fine-Grained Control for Song Generation

Pengfei Cai, Joanna Wang, Haorui Zheng, Xu Li, Zihao Ji, Teng Ma, Zhongliang Liu, Chen Zhang, Pengfei Wan

Recent advancements in song generation have shown promising results in generating songs from lyrics and/or global text prompts. However, most existing systems lack the ability to model the temporally varying attributes of songs, limiting fine-grained control over musical structure and dynamics. In this paper, we propose SegTune, a non-autoregressive framework for structured and controllable song generation. SegTune enables segment-level control by allowing users or large language models to specify local musical descriptions aligned to song sections.The segmental prompts are injected into the model by temporally broadcasting them to corresponding time windows, while global prompts influence the whole song to ensure stylistic coherence. To obtain accurate segment durations and enable precise lyric-to-music alignment, we introduce an LLM-based duration predictor that autoregressively generates sentence-level timestamped lyrics in LRC format. We further construct a large-scale data pipeline for collecting high-quality songs with aligned lyrics and prompts, and propose new evaluation metrics to assess segment-level alignment and vocal attribute consistency. Experimental results show that SegTune achieves superior controllability and musical coherence compared to existing baselines. See https://cai525.github.io/SegTune_demo for demos of our work.

☆ MVDR Beamforming for Cyclostationary Processes

Giovanni Bologni, Martin Bo Møller, Richard Heusdens, Richard C. Hendriks

Conventional acoustic beamformers assume that noise is stationary within short time frames. This assumption prevents them from exploiting correlations between frequencies in almost-periodic noise sources such as musical instruments, fans, and engines. These signals exhibit periodically varying statistics and are better modeled as cyclostationary processes. This paper introduces the cyclic MVDR (cMVDR) beamformer, an extension of the conventional MVDR that leverages both spatial and spectral correlations to improve noise reduction, particularly in low-SNR scenarios. The method builds on frequency-shifted (FRESH) filtering, where shifted versions of the input are combined to attenuate or amplify components that are coherent across frequency. To address inharmonicity, where harmonic partials deviate from exact integer multiples of the fundamental frequency, we propose a data-driven strategy that estimates resonant frequencies via periodogram analysis and computes the frequency shifts from their spacing. Analytical and experimental results demonstrate that performance improves with increasing spectral correlation. On real recordings, the cMVDR achieves up to 5 dB gain in scale-invariant signal-to-distortion ratio (SI-SDR) over the MVDR and remains effective even with a single microphone. Code is available at https://github.com/Screeen/cMVDR.

comment: Under review for publication from September 2025

☆ ParaStyleTTS: Toward Efficient and Robust Paralinguistic Style Control for Expressive Text-to-Speech Generation

Haowei Lou, Hye-Young Paik, Wen Hu, Lina Yao

Controlling speaking style in text-to-speech (TTS) systems has become a growing focus in both academia and industry. While many existing approaches rely on reference audio to guide style generation, such methods are often impractical due to privacy concerns and limited accessibility. More recently, large language models (LLMs) have been used to control speaking style through natural language prompts; however, their high computational cost, lack of interpretability, and sensitivity to prompt phrasing limit their applicability in real-time and resource-constrained environments. In this work, we propose ParaStyleTTS, a lightweight and interpretable TTS framework that enables expressive style control from text prompts alone. ParaStyleTTS features a novel two-level style adaptation architecture that separates prosodic and paralinguistic speech style modeling. It allows fine-grained and robust control over factors such as emotion, gender, and age. Unlike LLM-based methods, ParaStyleTTS maintains consistent style realization across varied prompt formulations and is well-suited for real-world applications, including on-device and low-resource deployment. Experimental results show that ParaStyleTTS generates high-quality speech with performance comparable to state-of-the-art LLM-based systems while being 30x faster, using 8x fewer parameters, and requiring 2.5x less CUDA memory. Moreover, ParaStyleTTS exhibits superior robustness and controllability over paralinguistic speaking styles, providing a practical and efficient solution for style-controllable text-to-speech generation. Demo can be found at https://parastyletts.github.io/ParaStyleTTS_Demo/. Code can be found at https://github.com/haoweilou/ParaStyleTTS.

☆ Adaptive Per-Channel Energy Normalization Front-end for Robust Audio Signal Processing ICASSP2026

Hanyu Meng, Vidhyasaharan Sethu, Eliathamby Ambikairajah, Qiquan Zhang, Haizhou Li

In audio signal processing, learnable front-ends have shown strong performance across diverse tasks by optimizing task-specific representation. However, their parameters remain fixed once trained, lacking flexibility during inference and limiting robustness under dynamic complex acoustic environments. In this paper, we introduce a novel adaptive paradigm for audio front-ends that replaces static parameterization with a closed-loop neural controller. Specifically, we simplify the learnable front-end LEAF architecture and integrate a neural controller for adaptive representation via dynamically tuning Per-Channel Energy Normalization. The neural controller leverages both the current and the buffered past subband energies to enable input-dependent adaptation during inference. Experimental results on multiple audio classification tasks demonstrate that the proposed adaptive front-end consistently outperforms prior fixed and learnable front-ends under both clean and complex acoustic conditions. These results highlight neural adaptability as a promising direction for the next generation of audio front-ends.

comment: Submitted to ICASSP2026

☆ Joint Estimation of Piano Dynamics and Metrical Structure with a Multi-task Multi-Scale Network ICASSP2026

Zhanhong He, Hanyu Meng, David Huang, Roberto Togneri

Estimating piano dynamic from audio recordings is a fundamental challenge in computational music analysis. In this paper, we propose an efficient multi-task network that jointly predicts dynamic levels, change points, beats, and downbeats from a shared latent representation. These four targets form the metrical structure of dynamics in the music score. Inspired by recent vocal dynamic research, we use a multi-scale network as the backbone, which takes Bark-scale specific loudness as the input feature. Compared to log-Mel as input, this reduces model size from 14.7 M to 0.5 M, enabling long sequential input. We use a 60-second audio length in audio segmentation, which doubled the length of beat tracking commonly used. Evaluated on the public MazurkaBL dataset, our model achieves state-of-the-art results across all tasks. This work sets a new benchmark for piano dynamic estimation and delivers a powerful and compact tool, paving the way for large-scale, resource-efficient analysis of musical expression.

comment: Paper submitted to ICASSP2026

☆ Diffusion Buffer for Online Generative Speech Enhancement

Bunlong Lay, Rostislav Makarov, Simon Welker, Maris Hillemann, Timo Gerkmann

Online Speech Enhancement was mainly reserved for predictive models. A key advantage of these models is that for an incoming signal frame from a stream of data, the model is called only once for enhancement. In contrast, generative Speech Enhancement models often require multiple calls, resulting in a computational complexity that is too high for many online speech enhancement applications. This work presents the Diffusion Buffer, a generative diffusion-based Speech Enhancement model which only requires one neural network call per incoming signal frame from a stream of data and performs enhancement in an online fashion on a consumer-grade GPU. The key idea of the Diffusion Buffer is to align physical time with Diffusion time-steps. The approach progressively denoises frames through physical time, where past frames have more noise removed. Consequently, an enhanced frame is output to the listener with a delay defined by the Diffusion Buffer, and the output frame has a corresponding look-ahead. In this work, we extend upon our previous work by carefully designing a 2D convolutional UNet architecture that specifically aligns with the Diffusion Buffer's look-ahead. We observe that the proposed UNet improves performance, particularly when the algorithmic latency is low. Moreover, we show that using a Data Prediction loss instead of Denoising Score Matching loss enables flexible control over the trade-off between algorithmic latency and quality during inference. The extended Diffusion Buffer equipped with a novel NN and loss function drastically reduces the algorithmic latency from 320 - 960 ms to 32 - 176 ms with an even increased performance. While it has been shown before that offline generative diffusion models outperform predictive approaches in unseen noisy speech data, we confirm that the online Diffusion Buffer also outperforms its predictive counterpart on unseen noisy speech data.

☆ Adapting Language Balance in Code-Switching Speech ICASSP 2026

Enes Yavuz Ugan, Ngoc-Quan Pham, Alexander Waibel

Despite achieving impressive results on standard benchmarks, large foundational models still struggle against code-switching test cases. When data scarcity cannot be used as the usual justification for poor performance, the reason may lie in the infrequent occurrence of code-switched moments, where the embedding of the second language appears subtly. Instead of expecting the models to learn this infrequency on their own, it might be beneficial to provide the training process with labels. Evaluating model performance on code-switching data requires careful localization of code-switching points where recognition errors are most consequential, so that the analysis emphasizes mistakes occurring at those moments. Building on this observation, we leverage the difference between the embedded and the main language to highlight those code-switching points and thereby emphasize learning at those locations. This simple yet effective differentiable surrogate mitigates context bias during generation -- the central challenge in code-switching -- thereby improving the model's robustness. Our experiments with Arabic and Chinese-English showed that the models are able to predict the switching places more correctly, reflected by the reduced substitution error.

comment: Submitted to ICASSP 2026

☆ Bayesian Low-Rank Factorization for Robust Model Adaptation ICASSP 2026

Enes Yavuz Ugan, Ngoc-Quan Pham, Alexander Waibel

Large speech foundation models achieve strong performance across many domains, but they often require adaptation to handle local needs such as code-switching, where speakers mix languages within the same utterance. Direct fine-tuning of these models risks overfitting to the target domain and overwriting the broad capabilities of the base model. To address this challenge, we explore Bayesian factorized adapters for speech foundation models, which place priors near zero to achieve sparser adaptation matrices and thereby retain general performance while adapting to specific domains. We apply our approach to the Whisper model and evaluate on different multilingual code-switching scenarios. Our results show only minimal adaptation loss while significantly reducing catastrophic forgetting of the base model. Compared to LoRA, our method achieves a backward gain of 54% with only a 4% drop on the new domain. These findings highlight the effectiveness of Bayesian adaptation for fine-tuning speech foundation models without sacrificing generalization.

comment: Submitted to ICASSP 2026

☆ Noise-Conditioned Mixture-of-Experts Framework for Robust Speaker Verification

Bin Gu, Lipeng Dai, Huipeng Du, Haitao Zhao, Jibo Wei

Robust speaker verification under noisy conditions remains an open challenge. Conventional deep learning methods learn a robust unified speaker representation space against diverse background noise and achieve significant improvement. In contrast, this paper presents a noise-conditioned mixture-ofexperts framework that decomposes the feature space into specialized noise-aware subspaces for speaker verification. Specifically, we propose a noise-conditioned expert routing mechanism, a universal model based expert specialization strategy, and an SNR-decaying curriculum learning protocol, collectively improving model robustness and generalization under diverse noise conditions. The proposed method can automatically route inputs to expert networks based on noise information derived from the inputs, where each expert targets distinct noise characteristics while preserving speaker identity information. Comprehensive experiments demonstrate consistent superiority over baselines, confirming that explicit noise-dependent feature modeling significantly enhances robustness without sacrificing verification accuracy.

☆ A Stage-Wise Learning Strategy with Fixed Anchors for Robust Speaker Verification

Bin Gu, Lipeng Dai, Huipeng Du, Haitao Zhao, Jibo Wei

Learning robust speaker representations under noisy conditions presents significant challenges, which requires careful handling of both discriminative and noise-invariant properties. In this work, we proposed an anchor-based stage-wise learning strategy for robust speaker representation learning. Specifically, our approach begins by training a base model to establish discriminative speaker boundaries, and then extract anchor embeddings from this model as stable references. Finally, a copy of the base model is fine-tuned on noisy inputs, regularized by enforcing proximity to their corresponding fixed anchor embeddings to preserve speaker identity under distortion. Experimental results suggest that this strategy offers advantages over conventional joint optimization, particularly in maintaining discrimination while improving noise robustness. The proposed method demonstrates consistent improvements across various noise conditions, potentially due to its ability to handle boundary stabilization and variation suppression separately.

☆ ProLAP: Probabilistic Language-Audio Pre-Training

Toranosuke Manabe, Yuchi Ishikawa, Hokuto Munakata, Tatsuya Komatsu

Language-audio joint representation learning frameworks typically depend on deterministic embeddings, assuming a one-to-one correspondence between audio and text. In real-world settings, however, the language-audio relationship is inherently many-to-many: one audio segment can be described by multiple captions and vice versa. To address this, we propose Probabilistic Language-Audio Pre-training (ProLAP), which models multiplicity as the spread of probability distributions in a joint language-audio embedding space. To train the intra-modal hierarchical relationship effectively, we also introduce two objectives: (i) hierarchical inclusion loss to promote semantic hierarchical understanding of inputs and (ii) mask repulsive loss to improve the efficiency of learning when optimizing the hierarchical inclusion loss. With this training strategy, our model can learn the hierarchical structure inherent in the data even from small datasets, in contrast to prior probabilistic approaches that rely on large-scale datasets. In our experiments, ProLAP outperforms existing deterministic approaches on audio-text retrieval tasks. Moreover, through experiments on the audio traversal task introduced in this paper, we demonstrate that ProLAP captures the plausible semantic hierarchy.

comment: Under review

☆ MVDR Beamforming for Cyclostationary Processes

Giovanni Bologni, Martin Bo Møller, Richard Heusdens, Richard C. Hendriks

Conventional acoustic beamformers assume that noise is stationary within short time frames. This assumption prevents them from exploiting correlations between frequencies in almost-periodic noise sources such as musical instruments, fans, and engines. These signals exhibit periodically varying statistics and are better modeled as cyclostationary processes. This paper introduces the cyclic MVDR (cMVDR) beamformer, an extension of the conventional MVDR that leverages both spatial and spectral correlations to improve noise reduction, particularly in low-SNR scenarios. The method builds on frequency-shifted (FRESH) filtering, where shifted versions of the input are combined to attenuate or amplify components that are coherent across frequency. To address inharmonicity, where harmonic partials deviate from exact integer multiples of the fundamental frequency, we propose a data-driven strategy that estimates resonant frequencies via periodogram analysis and computes the frequency shifts from their spacing. Analytical and experimental results demonstrate that performance improves with increasing spectral correlation. On real recordings, the cMVDR achieves up to 5 dB gain in scale-invariant signal-to-distortion ratio (SI-SDR) over the MVDR and remains effective even with a single microphone. Code is available at https://github.com/Screeen/cMVDR.

comment: Under review for publication from September 2025

☆ ParaStyleTTS: Toward Efficient and Robust Paralinguistic Style Control for Expressive Text-to-Speech Generation

Haowei Lou, Hye-Young Paik, Wen Hu, Lina Yao

Controlling speaking style in text-to-speech (TTS) systems has become a growing focus in both academia and industry. While many existing approaches rely on reference audio to guide style generation, such methods are often impractical due to privacy concerns and limited accessibility. More recently, large language models (LLMs) have been used to control speaking style through natural language prompts; however, their high computational cost, lack of interpretability, and sensitivity to prompt phrasing limit their applicability in real-time and resource-constrained environments. In this work, we propose ParaStyleTTS, a lightweight and interpretable TTS framework that enables expressive style control from text prompts alone. ParaStyleTTS features a novel two-level style adaptation architecture that separates prosodic and paralinguistic speech style modeling. It allows fine-grained and robust control over factors such as emotion, gender, and age. Unlike LLM-based methods, ParaStyleTTS maintains consistent style realization across varied prompt formulations and is well-suited for real-world applications, including on-device and low-resource deployment. Experimental results show that ParaStyleTTS generates high-quality speech with performance comparable to state-of-the-art LLM-based systems while being 30x faster, using 8x fewer parameters, and requiring 2.5x less CUDA memory. Moreover, ParaStyleTTS exhibits superior robustness and controllability over paralinguistic speaking styles, providing a practical and efficient solution for style-controllable text-to-speech generation. Demo can be found at https://parastyletts.github.io/ParaStyleTTS_Demo/. Code can be found at https://github.com/haoweilou/ParaStyleTTS.

☆ Adaptive Per-Channel Energy Normalization Front-end for Robust Audio Signal Processing ICASSP2026

Hanyu Meng, Vidhyasaharan Sethu, Eliathamby Ambikairajah, Qiquan Zhang, Haizhou Li

In audio signal processing, learnable front-ends have shown strong performance across diverse tasks by optimizing task-specific representation. However, their parameters remain fixed once trained, lacking flexibility during inference and limiting robustness under dynamic complex acoustic environments. In this paper, we introduce a novel adaptive paradigm for audio front-ends that replaces static parameterization with a closed-loop neural controller. Specifically, we simplify the learnable front-end LEAF architecture and integrate a neural controller for adaptive representation via dynamically tuning Per-Channel Energy Normalization. The neural controller leverages both the current and the buffered past subband energies to enable input-dependent adaptation during inference. Experimental results on multiple audio classification tasks demonstrate that the proposed adaptive front-end consistently outperforms prior fixed and learnable front-ends under both clean and complex acoustic conditions. These results highlight neural adaptability as a promising direction for the next generation of audio front-ends.

comment: Submitted to ICASSP2026

☆ Joint Estimation of Piano Dynamics and Metrical Structure with a Multi-task Multi-Scale Network ICASSP2026

Zhanhong He, Hanyu Meng, David Huang, Roberto Togneri

Estimating piano dynamic from audio recordings is a fundamental challenge in computational music analysis. In this paper, we propose an efficient multi-task network that jointly predicts dynamic levels, change points, beats, and downbeats from a shared latent representation. These four targets form the metrical structure of dynamics in the music score. Inspired by recent vocal dynamic research, we use a multi-scale network as the backbone, which takes Bark-scale specific loudness as the input feature. Compared to log-Mel as input, this reduces model size from 14.7 M to 0.5 M, enabling long sequential input. We use a 60-second audio length in audio segmentation, which doubled the length of beat tracking commonly used. Evaluated on the public MazurkaBL dataset, our model achieves state-of-the-art results across all tasks. This work sets a new benchmark for piano dynamic estimation and delivers a powerful and compact tool, paving the way for large-scale, resource-efficient analysis of musical expression.

comment: Paper submitted to ICASSP2026

♻ ☆ Lightweight and Robust Multi-Channel End-to-End Speech Recognition with Spherical Harmonic Transform

Xiangzhu Kong, Huang Hao, Zhijian Ou

This paper presents SHTNet, a lightweight spherical harmonic transform (SHT) based framework, which is designed to address cross-array generalization challenges in multi-channel automatic speech recognition (ASR) through three key innovations. First, SHT based spatial sound field decomposition converts microphone signals into geometry-invariant spherical harmonic coefficients, isolating signal processing from array geometry. Second, the Spatio-Spectral Attention Fusion Network (SSAFN) combines coordinate-aware spatial modeling, refined self-attention channel combinator, and spectral noise suppression without conventional beamforming. Third, Rand-SHT training enhances robustness through random channel selection and array geometry reconstruction. The system achieves 39.26\% average CER across heterogeneous arrays (e.g., circular, square, and binaural) on datasets including Aishell-4, Alimeeting, and XMOS, with 97.1\% fewer computations than conventional neural beamformers.

comment: Interspeech 2025

♻ ☆ Post-training for Deepfake Speech Detection

Wanying Ge, Xin Wang, Xuechen Liu, Junichi Yamagishi

We introduce a post-training approach that adapts self-supervised learning (SSL) models for deepfake speech detection by bridging the gap between general pre-training and domain-specific fine-tuning. We present AntiDeepfake models, a series of post-trained models developed using a large-scale multilingual speech dataset containing over 56,000 hours of genuine speech and 18,000 hours of speech with various artifacts in over one hundred languages. Experimental results show that the post-trained models already exhibit strong robustness and generalization to unseen deepfake speech. When they are further fine-tuned on the Deepfake-Eval-2024 dataset, these models consistently surpass existing state-of-the-art detectors that do not leverage post-training. Model checkpoints and source code are available online.

comment: Corrected previous implementation of EER calculation. Slight numerical changes in some of the results

☆ DELULU: Discriminative Embedding Learning Using Latent Units for Speaker-Aware Self-Supervised Speech Foundational Model

Massa Baali, Rita Singh, Bhiksha Raj

Self-supervised speech models have achieved remarkable success on content-driven tasks, yet they remain limited in capturing speaker-discriminative features critical for verification, diarization, and profiling applications. We introduce DELULU, a speaker-aware self-supervised foundational model that addresses this limitation by integrating external supervision into the pseudo-label generation process. DELULU leverages frame-level embeddings from ReDimNet, a state-of-the-art speaker verification model, to guide the k-means clustering step during pre-training, introducing a strong speaker-discriminative inductive bias that aligns representation learning with speaker identity. The model is trained using a dual objective that combines masked prediction and denoising, further enhancing robustness and generalization. DELULU significantly outperforms prior self-supervised learning (SSL) models across a range of speaker-centric tasks, achieving up to 62% relative improvement in equal error rate (EER) for speaker verification and consistent gains on zero-shot profiling tasks such as gender, age, accent, and speaker counting. Our findings demonstrate that DELULU is a strong universal encoder for speaker-aware speech processing, enabling superior performance even without task-specific fine-tuning.

☆ SARSteer: Safeguarding Large Audio Language Models via Safe-Ablated Refusal Steering

Weilin Lin, Jianze Li, Hui Xiong, Li Liu

Large Audio-Language Models (LALMs) are becoming essential as a powerful multimodal backbone for real-world applications. However, recent studies show that audio inputs can more easily elicit harmful responses than text, exposing new risks toward deployment. While safety alignment has made initial advances in LLMs and Large Vision-Language Models (LVLMs), we find that vanilla adaptation of these approaches to LALMs faces two key limitations: 1) LLM-based steering fails under audio input due to the large distributional gap between activations, and 2) prompt-based defenses induce over-refusals on benign-speech queries. To address these challenges, we propose Safe-Ablated Refusal Steering (SARSteer), the first inference-time defense framework for LALMs. Specifically, SARSteer leverages text-derived refusal steering to enforce rejection without manipulating audio inputs and introduces decomposed safe-space ablation to mitigate over-refusal. Extensive experiments demonstrate that SARSteer significantly improves harmful-query refusal while preserving benign responses, establishing a principled step toward safety alignment in LALMs.

☆ AWARE: Audio Watermarking with Adversarial Resistance to Edits

Kosta Pavlović, Lazar Stanarević, Petar Nedić, Slavko Kovačević, Igor Djurović

Prevailing practice in learning-based audio watermarking is to pursue robustness by expanding the set of simulated distortions during training. However, such surrogates are narrow and prone to overfitting. This paper presents AWARE (Audio Watermarking with Adversarial Resistance to Edits), an alternative approach that avoids reliance on attack-simulation stacks and handcrafted differentiable distortions. Embedding is obtained via adversarial optimization in the time-frequency domain under a level-proportional perceptual budget. Detection employs a time-order-agnostic detector with a Bitwise Readout Head (BRH) that aggregates temporal evidence into one score per watermark bit, enabling reliable watermark decoding even under desynchronization and temporal cuts. Empirically, AWARE attains high audio quality and speech intelligibility (PESQ/STOI) and consistently low BER across various audio edits, often surpassing representative state-of-the-art learning-based audio watermarking systems.

☆ Not All Deepfakes Are Created Equal: Triaging Audio Forgeries for Robust Deepfake Singer Identification NeurIPS 2025

Davide Salvi, Hendrik Vincent Koops, Elio Quinton

The proliferation of highly realistic singing voice deepfakes presents a significant challenge to protecting artist likeness and content authenticity. Automatic singer identification in vocal deepfakes is a promising avenue for artists and rights holders to defend against unauthorized use of their voice, but remains an open research problem. Based on the premise that the most harmful deepfakes are those of the highest quality, we introduce a two-stage pipeline to identify a singer's vocal likeness. It first employs a discriminator model to filter out low-quality forgeries that fail to accurately reproduce vocal likeness. A subsequent model, trained exclusively on authentic recordings, identifies the singer in the remaining high-quality deepfakes and authentic audio. Experiments show that this system consistently outperforms existing baselines on both authentic and synthetic content.

comment: Accepted for presentation at the NeurIPS 2025 Workshop on Generative and Protective AI for Content Creation (non-archival)

☆ TopSeg: A Multi-Scale Topological Framework for Data-Efficient Heart Sound Segmentation ICASSP2026

Peihong Zhang, Zhixin Li, Yuxuan Liu, Rui Sang, Yiqiang Cai, Yizhou Tan, Shengchen Li

Deep learning approaches for heart-sound (PCG) segmentation built on time--frequency features can be accurate but often rely on large expert-labeled datasets, limiting robustness and deployment. We present TopSeg, a topological representation-centric framework that encodes PCG dynamics with multi-scale topological features and decodes them using a lightweight temporal convolutional network (TCN) with an order- and duration-constrained inference step. To evaluate data efficiency and generalization, we train exclusively on PhysioNet 2016 dataset with subject-level subsampling and perform external validation on CirCor dataset. Under matched-capacity decoders, the topological features consistently outperform spectrogram and envelope inputs, with the largest margins at low data budgets; as a full system, TopSeg surpasses representative end-to-end baselines trained on their native inputs under the same budgets while remaining competitive at full data. Ablations at 10% training confirm that all scales contribute and that combining H_0 and H_1 yields more reliable S1/S2 localization and boundary stability. These results indicate that topology-aware representations provide a strong inductive bias for data-efficient, cross-dataset PCG segmentation, supporting practical use when labeled data are limited.

comment: Paper has submitted to ICASSP2026

☆ DDSC: Dynamic Dual-Signal Curriculum for Data-Efficient Acoustic Scene Classification under Domain Shift ICASSP2026

Peihong Zhang, Yuxuan Liu, Rui Sang, Zhixin Li, Yiqiang Cai, Yizhou Tan, Shengchen Li

Acoustic scene classification (ASC) suffers from device-induced domain shift, especially when labels are limited. Prior work focuses on curriculum-based training schedules that structure data presentation by ordering or reweighting training examples from easy-to-hard to facilitate learning; however, existing curricula are static, fixing the ordering or the weights before training and ignoring that example difficulty and marginal utility evolve with the learned representation. To overcome this limitation, we propose the Dynamic Dual-Signal Curriculum (DDSC), a training schedule that adapts the curriculum online by combining two signals computed each epoch: a domain-invariance signal and a learning-progress signal. A time-varying scheduler fuses these signals into per-example weights that prioritize domain-invariant examples in early epochs and progressively emphasize device-specific cases. DDSC is lightweight, architecture-agnostic, and introduces no additional inference overhead. Under the official DCASE 2024 Task~1 protocol, DDSC consistently improves cross-device performance across diverse ASC baselines and label budgets, with the largest gains on unseen-device splits.

comment: Paper has submitted to ICASSP2026

☆ Event Topology-based Visual Microphone for Amplitude and Frequency Reconstruction

Ryogo Niwa, Yoichi Ochiai, Tatsuki Fushimi

Accurate vibration measurement is vital for analyzing dynamic systems across science and engineering, yet noncontact methods often balance precision against practicality. Event cameras offer high-speed, low-light sensing, but existing approaches fail to recover vibration amplitude and frequency with sufficient accuracy. We present an event topology-based visual microphone that reconstructs vibrations directly from raw event streams without external illumination. By integrating the Mapper algorithm from topological data analysis with hierarchical density-based clustering, our framework captures the intrinsic structure of event data to recover both amplitude and frequency with high fidelity. Experiments demonstrate substantial improvements over prior methods and enable simultaneous recovery of multiple sound sources from a single event stream, advancing the frontier of passive, illumination-free vibration sensing.

comment: 6 pages, 5 figures, 2 tables. Submitted for publication

☆ Hearing Health in Home Healthcare: Leveraging LLMs for Illness Scoring and ALMs for Vocal Biomarker Extraction NeurIPS 2025

Yu-Wen Chen, William Ho, Sasha M. Vergez, Grace Flaherty, Pallavi Gupta, Zhihong Zhang, Maryam Zolnoori, Margaret V. McDonald, Maxim Topaz, Zoran Kostic, Julia Hirschberg

The growing demand for home healthcare calls for tools that can support care delivery. In this study, we explore automatic health assessment from voice using real-world home care visit data, leveraging the diverse patient information it contains. First, we utilize Large Language Models (LLMs) to integrate Subjective, Objective, Assessment, and Plan (SOAP) notes derived from unstructured audio transcripts and structured vital signs into a holistic illness score that reflects a patient's overall health. This compact representation facilitates cross-visit health status comparisons and downstream analysis. Next, we design a multi-stage preprocessing pipeline to extract short speech segments from target speakers in home care recordings for acoustic analysis. We then employ an Audio Language Model (ALM) to produce plain-language descriptions of vocal biomarkers and examine their association with individuals' health status. Our experimental results benchmark both commercial and open-source LLMs in estimating illness scores, demonstrating their alignment with actual clinical outcomes, and revealing that SOAP notes are substantially more informative than vital signs. Building on the illness scores, we provide the first evidence that ALMs can identify health-related acoustic patterns from home care recordings and present them in a human-readable form. Together, these findings highlight the potential of LLMs and ALMs to harness heterogeneous in-home visit data for better patient monitoring and care.

comment: The Second Workshop on GenAI for Health at NeurIPS 2025

☆ Transformer Redesign for Late Fusion of Audio-Text Features on Ultra-Low-Power Edge Hardware

Stavros Mitsis, Ermos Hadjikyriakos, Humaid Ibrahim, Savvas Neofytou, Shashwat Raman, James Myles, Eiman Kanjo

Deploying emotion recognition systems in real-world environments where devices must be small, low-power, and private remains a significant challenge. This is especially relevant for applications such as tension monitoring, conflict de-escalation, and responsive wearables, where cloud-based solutions are impractical. Multimodal emotion recognition has advanced through deep learning, but most systems remain unsuitable for deployment on ultra-constrained edge devices. Prior work typically relies on powerful hardware, lacks real-time performance, or uses unimodal input. This paper addresses that gap by presenting a hardware-aware emotion recognition system that combines acoustic and linguistic features using a late-fusion architecture optimised for Edge TPU. The design integrates a quantised transformer-based acoustic model with frozen keyword embeddings from a DSResNet-SE network, enabling real-time inference within a 1.8MB memory budget and 21-23ms latency. The pipeline ensures spectrogram alignment between training and deployment using MicroFrontend and MLTK. Evaluation on re-recorded, segmented IEMOCAP samples captured through the Coral Dev Board Micro microphone shows a 6.3% macro F1 improvement over unimodal baselines. This work demonstrates that accurate, real-time multimodal emotion inference is achievable on microcontroller-class edge platforms through task-specific fusion and hardware-guided model design.

♻ ☆ Test-Time Training for Speech Enhancement

Avishkar Behera, Riya Ann Easow, Venkatesh Parvathala, K. Sri Rama Murty

This paper introduces a novel application of Test-Time Training (TTT) for Speech Enhancement, addressing the challenges posed by unpredictable noise conditions and domain shifts. This method combines a main speech enhancement task with a self-supervised auxiliary task in a Y-shaped architecture. The model dynamically adapts to new domains during inference time by optimizing the proposed self-supervised tasks like noise-augmented signal reconstruction or masked spectrogram prediction, bypassing the need for labeled data. We further introduce various TTT strategies offering a trade-off between adaptation and efficiency. Evaluations across synthetic and real-world datasets show consistent improvements across speech quality metrics, outperforming the baseline model. This work highlights the effectiveness of TTT in speech enhancement, providing insights for future research in adaptive and robust speech processing.

comment: Published in the Proceedings of Interspeech 2025

♻ ☆ Nexus: An Omni-Perceptive And -Interactive Model for Language, Audio, And Vision

Che Liu, Yingji Zhang, Dong Zhang, Weijie Zhang, Chenggong Gong, Yu Lu, Shilin Zhou, Ziliang Gan, Ziao Wang, Haipang Wu, Ji Liu, André Freitas, Qifan Wang, Zenglin Xu, Rongjuncheng Zhang, Yong Dai

This work proposes an industry-level omni-modal large language model (LLM) pipeline that integrates auditory, visual, and linguistic modalities to overcome challenges such as limited tri-modal datasets, high computational costs, and complex feature alignments. Our pipeline consists of three main components: First, a modular framework enabling flexible configuration of various encoder-LLM-decoder architectures. Second, a lightweight training strategy that pre-trains audio-language alignment on the state-of-the-art vision-language model Qwen2.5-VL, thus avoiding the costly pre-training of vision-specific modalities. Third, an audio synthesis pipeline that generates high-quality audio-text data from diverse real-world scenarios, supporting applications such as Automatic Speech Recognition and Speech-to-Speech chat. To this end, we introduce an industry-level omni-modal LLM, Nexus. Extensive experiments validate the efficacy of our pipeline, yielding the following key findings:(1) In the visual understanding task, Nexus exhibits superior performance compared with its backbone model - Qwen2.5-VL-7B, validating the efficiency of our training strategy. (2) Within the English Spoken Question-Answering task, the model achieves better accuracy than the same-period competitor (i.e, MiniCPM-o2.6-7B) in the LLaMA Q. benchmark. (3) In our real-world ASR testset, Nexus achieves outstanding performance, indicating its robustness in real scenarios. (4) In the Speech-to-Text Translation task, our model outperforms Qwen2-Audio-Instruct-7B. (5) In the Text-to-Speech task, based on pretrained vocoder (e.g., Fishspeech1.4 or CosyVoice2.0), Nexus is comparable to its backbone vocoder on Seed-TTS benchmark. (6) An in-depth analysis of tri-modal alignment reveals that incorporating the audio modality enhances representational alignment between vision and language.

comment: Project: https://github.com/HiThink-Research/NEXUS-O

♻ ☆ 3D Audio-Visual Segmentation NeurIPS 2024

Artem Sokolov, Swapnil Bhosale, Xiatian Zhu

Recognizing the sounding objects in scenes is a longstanding objective in embodied AI, with diverse applications in robotics and AR/VR/MR. To that end, Audio-Visual Segmentation (AVS), taking as condition an audio signal to identify the masks of the target sounding objects in an input image with synchronous camera and microphone sensors, has been recently advanced. However, this paradigm is still insufficient for real-world operation, as the mapping from 2D images to 3D scenes is missing. To address this fundamental limitation, we introduce a novel research problem, 3D Audio-Visual Segmentation, extending the existing AVS to the 3D output space. This problem poses more challenges due to variations in camera extrinsics, audio scattering, occlusions, and diverse acoustics across sounding object categories. To facilitate this research, we create the very first simulation based benchmark, 3DAVS-S34-O7, providing photorealistic 3D scene environments with grounded spatial audio under single-instance and multi-instance settings, across 34 scenes and 7 object categories. This is made possible by re-purposing the Habitat simulator to generate comprehensive annotations of sounding object locations and corresponding 3D masks. Subsequently, we propose a new approach, EchoSegnet, characterized by integrating the ready-to-use knowledge from pretrained 2D audio-visual foundation models synergistically with 3D visual scene representation through spatial audio-aware mask alignment and refinement. Extensive experiments demonstrate that EchoSegnet can effectively segment sounding objects in 3D space on our new benchmark, representing a significant advancement in the field of embodied AI. Project page: https://x-up-lab.github.io/research/3d-audio-visual-segmentation/

comment: Accepted at the NeurIPS 2024 Workshop on Audio Imagination; this version updates the project page link

☆ AnyRIR: Robust Non-intrusive Room Impulse Response Estimation in the Wild

Kyung Yun Lee, Nils Meyer-Kahlen, Karolina Prawda, Vesa Välimäki, Sebastian J. Schlecht

We address the problem of estimating room impulse responses (RIRs) in noisy, uncontrolled environments where non-stationary sounds such as speech or footsteps corrupt conventional deconvolution. We propose AnyRIR, a non-intrusive method that uses music as the excitation signal instead of a dedicated test signal, and formulate RIR estimation as an L1-norm regression in the time-frequency domain. Solved efficiently with Iterative Reweighted Least Squares (IRLS) and Least-Squares Minimal Residual (LSMR) methods, this approach exploits the sparsity of non-stationary noise to suppress its influence. Experiments on simulated and measured data show that AnyRIR outperforms L2-based and frequency-domain deconvolution, under in-the-wild noisy scenarios and codec mismatch, enabling robust RIR estimation for AR/VR and related applications.

☆ Hearing Health in Home Healthcare: Leveraging LLMs for Illness Scoring and ALMs for Vocal Biomarker Extraction NeurIPS 2025

Yu-Wen Chen, William Ho, Sasha M. Vergez, Grace Flaherty, Pallavi Gupta, Zhihong Zhang, Maryam Zolnoori, Margaret V. McDonald, Maxim Topaz, Zoran Kostic, Julia Hirschberg

The growing demand for home healthcare calls for tools that can support care delivery. In this study, we explore automatic health assessment from voice using real-world home care visit data, leveraging the diverse patient information it contains. First, we utilize Large Language Models (LLMs) to integrate Subjective, Objective, Assessment, and Plan (SOAP) notes derived from unstructured audio transcripts and structured vital signs into a holistic illness score that reflects a patient's overall health. This compact representation facilitates cross-visit health status comparisons and downstream analysis. Next, we design a multi-stage preprocessing pipeline to extract short speech segments from target speakers in home care recordings for acoustic analysis. We then employ an Audio Language Model (ALM) to produce plain-language descriptions of vocal biomarkers and examine their association with individuals' health status. Our experimental results benchmark both commercial and open-source LLMs in estimating illness scores, demonstrating their alignment with actual clinical outcomes, and revealing that SOAP notes are substantially more informative than vital signs. Building on the illness scores, we provide the first evidence that ALMs can identify health-related acoustic patterns from home care recordings and present them in a human-readable form. Together, these findings highlight the potential of LLMs and ALMs to harness heterogeneous in-home visit data for better patient monitoring and care.

comment: The Second Workshop on GenAI for Health at NeurIPS 2025

☆ Transformer Redesign for Late Fusion of Audio-Text Features on Ultra-Low-Power Edge Hardware

Stavros Mitsis, Ermos Hadjikyriakos, Humaid Ibrahim, Savvas Neofytou, Shashwat Raman, James Myles, Eiman Kanjo

Deploying emotion recognition systems in real-world environments where devices must be small, low-power, and private remains a significant challenge. This is especially relevant for applications such as tension monitoring, conflict de-escalation, and responsive wearables, where cloud-based solutions are impractical. Multimodal emotion recognition has advanced through deep learning, but most systems remain unsuitable for deployment on ultra-constrained edge devices. Prior work typically relies on powerful hardware, lacks real-time performance, or uses unimodal input. This paper addresses that gap by presenting a hardware-aware emotion recognition system that combines acoustic and linguistic features using a late-fusion architecture optimised for Edge TPU. The design integrates a quantised transformer-based acoustic model with frozen keyword embeddings from a DSResNet-SE network, enabling real-time inference within a 1.8MB memory budget and 21-23ms latency. The pipeline ensures spectrogram alignment between training and deployment using MicroFrontend and MLTK. Evaluation on re-recorded, segmented IEMOCAP samples captured through the Coral Dev Board Micro microphone shows a 6.3% macro F1 improvement over unimodal baselines. This work demonstrates that accurate, real-time multimodal emotion inference is achievable on microcontroller-class edge platforms through task-specific fusion and hardware-guided model design.

☆ AWARE: Audio Watermarking with Adversarial Resistance to Edits

Kosta Pavlović, Lazar Stanarević, Petar Nedić, Slavko Kovačević, Igor Djurović

Prevailing practice in learning-based audio watermarking is to pursue robustness by expanding the set of simulated distortions during training. However, such surrogates are narrow and prone to overfitting. This paper presents AWARE (Audio Watermarking with Adversarial Resistance to Edits), an alternative approach that avoids reliance on attack-simulation stacks and handcrafted differentiable distortions. Embedding is obtained via adversarial optimization in the time-frequency domain under a level-proportional perceptual budget. Detection employs a time-order-agnostic detector with a Bitwise Readout Head (BRH) that aggregates temporal evidence into one score per watermark bit, enabling reliable watermark decoding even under desynchronization and temporal cuts. Empirically, AWARE attains high audio quality and speech intelligibility (PESQ/STOI) and consistently low BER across various audio edits, often surpassing representative state-of-the-art learning-based audio watermarking systems.

♻ ☆ Speech Foundation Models Generalize to Time Series Tasks from Wearable Sensor Data

Jaya Narain, Zakaria Aldeneh, Shirley Ren

Both speech and sensor time series data encode information in both the time- and frequency- domains, like spectral powers and waveform shapelets. We show that speech foundation models learn representations that generalize beyond the speech domain and achieve state-of-the-art performance on diverse time-series tasks from wearable sensors. Probes trained on features extracted from HuBERT and wav2vec 2.0 outperform those extracted from self-supervised models trained directly on modality-specific datasets for mood classification, arrhythmia detection, and activity classification tasks. We find that the convolutional feature encoders of speech models are particularly relevant for wearable sensor applications. The proposed approach enhances performance on data-scarce time-series tasks using simple probing methods. This work takes a step toward developing generalized time-series models that unify speech and sensor modalities.

comment: Preprint, under review

♻ ☆ Test-Time Training for Speech Enhancement

Avishkar Behera, Riya Ann Easow, Venkatesh Parvathala, K. Sri Rama Murty

This paper introduces a novel application of Test-Time Training (TTT) for Speech Enhancement, addressing the challenges posed by unpredictable noise conditions and domain shifts. This method combines a main speech enhancement task with a self-supervised auxiliary task in a Y-shaped architecture. The model dynamically adapts to new domains during inference time by optimizing the proposed self-supervised tasks like noise-augmented signal reconstruction or masked spectrogram prediction, bypassing the need for labeled data. We further introduce various TTT strategies offering a trade-off between adaptation and efficiency. Evaluations across synthetic and real-world datasets show consistent improvements across speech quality metrics, outperforming the baseline model. This work highlights the effectiveness of TTT in speech enhancement, providing insights for future research in adaptive and robust speech processing.

comment: Published in the Proceedings of Interspeech 2025

♻ ☆ Nexus: An Omni-Perceptive And -Interactive Model for Language, Audio, And Vision

Che Liu, Yingji Zhang, Dong Zhang, Weijie Zhang, Chenggong Gong, Yu Lu, Shilin Zhou, Ziliang Gan, Ziao Wang, Haipang Wu, Ji Liu, André Freitas, Qifan Wang, Zenglin Xu, Rongjuncheng Zhang, Yong Dai

This work proposes an industry-level omni-modal large language model (LLM) pipeline that integrates auditory, visual, and linguistic modalities to overcome challenges such as limited tri-modal datasets, high computational costs, and complex feature alignments. Our pipeline consists of three main components: First, a modular framework enabling flexible configuration of various encoder-LLM-decoder architectures. Second, a lightweight training strategy that pre-trains audio-language alignment on the state-of-the-art vision-language model Qwen2.5-VL, thus avoiding the costly pre-training of vision-specific modalities. Third, an audio synthesis pipeline that generates high-quality audio-text data from diverse real-world scenarios, supporting applications such as Automatic Speech Recognition and Speech-to-Speech chat. To this end, we introduce an industry-level omni-modal LLM, Nexus. Extensive experiments validate the efficacy of our pipeline, yielding the following key findings:(1) In the visual understanding task, Nexus exhibits superior performance compared with its backbone model - Qwen2.5-VL-7B, validating the efficiency of our training strategy. (2) Within the English Spoken Question-Answering task, the model achieves better accuracy than the same-period competitor (i.e, MiniCPM-o2.6-7B) in the LLaMA Q. benchmark. (3) In our real-world ASR testset, Nexus achieves outstanding performance, indicating its robustness in real scenarios. (4) In the Speech-to-Text Translation task, our model outperforms Qwen2-Audio-Instruct-7B. (5) In the Text-to-Speech task, based on pretrained vocoder (e.g., Fishspeech1.4 or CosyVoice2.0), Nexus is comparable to its backbone vocoder on Seed-TTS benchmark. (6) An in-depth analysis of tri-modal alignment reveals that incorporating the audio modality enhances representational alignment between vision and language.

comment: Project: https://github.com/HiThink-Research/NEXUS-O

♻ ☆ Late Fusion and Multi-Level Fission Amplify Cross-Modal Transfer in Text-Speech LMs

Santiago Cuervo, Adel Moumen, Yanis Labrak, Sameer Khurana, Antoine Laurent, Mickael Rouvier, Phil Woodland, Ricard Marxer

Text-Speech Language Models (TSLMs) -- language models trained to jointly process and generate text and speech -- are commonly trained through an early modality fusion/fission approach, in which both modalities are fed and predicted from a shared backbone via linear layers. We hypothesize that this approach limits cross-modal transfer by neglecting feature compositionality -- specifically, the finer-grained nature of speech representations compared to text -- preventing the emergence of a shared feature hierarchy within model layers. In this paper, we argue that this limitation can be addressed through late fusion and fission, with a fission process that accesses both high- and low-level features for speech generation. Our models implementing these principles, SmolTolk, rival or surpass state-of-the-art TSLMs trained with orders of magnitude more compute, and achieve significantly improved cross-modal performance relative to early fusion/fission baselines. Representation analyses further suggest that our method enhances the model's ability to abstract higher-level, more semantic features from speech, and leads to increasingly shared representation spaces across layers.

♻ ☆ Post-training for Deepfake Speech Detection

Wanying Ge, Xin Wang, Xuechen Liu, Junichi Yamagishi

We introduce a post-training approach that adapts self-supervised learning (SSL) models for deepfake speech detection by bridging the gap between general pre-training and domain-specific fine-tuning. We present AntiDeepfake models, a series of post-trained models developed using a large-scale multilingual speech dataset containing over 56,000 hours of genuine speech and 18,000 hours of speech with various artifacts in over one hundred languages. Experimental results show that the post-trained models already exhibit strong robustness and generalization to unseen deepfake speech. When they are further fine-tuned on the Deepfake-Eval-2024 dataset, these models consistently surpass existing state-of-the-art detectors that do not leverage post-training. Model checkpoints and source code are available online.

comment: Corrected previous implementation of EER calculation. Slight numerical changes in some of the results

♻ ☆ 3D Audio-Visual Segmentation NeurIPS 2024

Artem Sokolov, Swapnil Bhosale, Xiatian Zhu

Recognizing the sounding objects in scenes is a longstanding objective in embodied AI, with diverse applications in robotics and AR/VR/MR. To that end, Audio-Visual Segmentation (AVS), taking as condition an audio signal to identify the masks of the target sounding objects in an input image with synchronous camera and microphone sensors, has been recently advanced. However, this paradigm is still insufficient for real-world operation, as the mapping from 2D images to 3D scenes is missing. To address this fundamental limitation, we introduce a novel research problem, 3D Audio-Visual Segmentation, extending the existing AVS to the 3D output space. This problem poses more challenges due to variations in camera extrinsics, audio scattering, occlusions, and diverse acoustics across sounding object categories. To facilitate this research, we create the very first simulation based benchmark, 3DAVS-S34-O7, providing photorealistic 3D scene environments with grounded spatial audio under single-instance and multi-instance settings, across 34 scenes and 7 object categories. This is made possible by re-purposing the Habitat simulator to generate comprehensive annotations of sounding object locations and corresponding 3D masks. Subsequently, we propose a new approach, EchoSegnet, characterized by integrating the ready-to-use knowledge from pretrained 2D audio-visual foundation models synergistically with 3D visual scene representation through spatial audio-aware mask alignment and refinement. Extensive experiments demonstrate that EchoSegnet can effectively segment sounding objects in 3D space on our new benchmark, representing a significant advancement in the field of embodied AI. Project page: https://x-up-lab.github.io/research/3d-audio-visual-segmentation/

comment: Accepted at the NeurIPS 2024 Workshop on Audio Imagination; this version updates the project page link

☆ SAKE: Towards Editing Auditory Attribute Knowledge of Large Audio-Language Models

Chih-Kai Yang, Yen-Ting Piao, Tzu-Wen Hsu, Szu-Wei Fu, Zhehuai Chen, Ke-Han Lu, Sung-Feng Huang, Chao-Han Huck Yang, Yu-Chiang Frank Wang, Yun-Nung Chen, Hung-yi Lee

Knowledge editing offers an efficient way to update model knowledge without full retraining, but prior work has concentrated almost exclusively on textual or visual modalities. We introduce SAKE, the first benchmark specifically designed for editing auditory attribute knowledge in Large Audio-Language Models (LALMs). Unlike factual updates, SAKE targets several abstract auditory attributes, capturing knowledge types that go beyond conventional textual and visual domains. We benchmark seven editing methods on two LALMs along four dimensions: reliability, generality, audio/text locality, and portability. Results highlight challenges such as preserving intra-attribute knowledge unrelated to the edit, generalizing edits to multimodal reasoning, and maintaining edits under sequential updates. SAKE provides a principled framework to study how knowledge editing extends to the auditory modalities, opening new directions for maintaining and adapting LALMs in more diverse real-world scenarios.

comment: Work in progress

☆ Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations ICASSP 2026

Bo-Han Feng, Chien-Feng Liu, Yu-Hsuan Li Liang, Chih-Kai Yang, Szu-Wei Fu, Zhehuai Chen, Ke-Han Lu, Sung-Feng Huang, Chao-Han Huck Yang, Yu-Chiang Frank Wang, Yun-Nung Chen, Hung-yi Lee

Large audio-language models (LALMs) extend text-based LLMs with auditory understanding, offering new opportunities for multimodal applications. While their perception, reasoning, and task performance have been widely studied, their safety alignment under paralinguistic variation remains underexplored. This work systematically investigates the role of speaker emotion. We construct a dataset of malicious speech instructions expressed across multiple emotions and intensities, and evaluate several state-of-the-art LALMs. Our results reveal substantial safety inconsistencies: different emotions elicit varying levels of unsafe responses, and the effect of intensity is non-monotonic, with medium expressions often posing the greatest risk. These findings highlight an overlooked vulnerability in LALMs and call for alignment strategies explicitly designed to ensure robustness under emotional variation, a prerequisite for trustworthy deployment in real-world settings.

comment: Submitted to ICASSP 2026

☆ Schrödinger Bridge Mamba for One-Step Speech Enhancement

Jing Yang, Sirui Wang, Chao Wu, Fan Fan

We propose Schr\"odinger Bridge Mamba (SBM), a new concept of training-inference framework motivated by the inherent compatibility between Schr\"odinger Bridge (SB) training paradigm and selective state-space model Mamba. We exemplify the concept of SBM with an implementation for generative speech enhancement. Experiments on a joint denoising and dereverberation task using four benchmark datasets demonstrate that SBM, with only 1-step inference, outperforms strong baselines with 1-step or iterative inference and achieves the best real-time factor (RTF). Beyond speech enhancement, we discuss the integration of SB paradigm and selective state-space model architecture based on their underlying alignment, which indicates a promising direction for exploring new deep generative models potentially applicable to a broad range of generative tasks. Demo page: https://sbmse.github.io

comment: 5 pages, 1 figure

☆ U-Codec: Ultra Low Frame-rate Neural Speech Codec for Fast High-fidelity Speech Generation

Xusheng Yang, Long Zhou, Wenfu Wang, Kai Hu, Shulin Feng, Chenxing Li, Meng Yu, Dong Yu, Yuexian Zou

We propose \textbf{U-Codec}, an \textbf{U}ltra low frame-rate neural speech \textbf{Codec} that achieves high-fidelity reconstruction and fast speech generation at an extremely low frame-rate of 5Hz (5 frames per second). Extreme compression at 5Hz typically leads to severe intelligibility and spectral detail loss, we introduce a Transformer-based inter-frame long-term dependency module and systematically explore residual vector quantization (RVQ) depth and codebook size to identify optimal configurations. Moreover, we apply U-Codec into a large language model (LLM)-based auto-regressive TTS model, which leverages global and local hierarchical architecture to effectively capture dependencies across multi-layer tokens. We extend LLM-based TTS from 3-layer RVQ at 50Hz to 32-layer RVQ at 5Hz. Experimental results demonstrate that U-Codec improves LLM-based TTS inference speed by around 3 $\times$ over high-frame-rate codecs while maintaining similarity and naturalness. These results validate the feasibility of using highly compressed 5Hz discrete tokens for fast and high-fidelity speech synthesis.

☆ Zero- and One-Shot Data Augmentation for Sentence-Level Dysarthric Speech Recognition in Constrained Scenarios SC 2025

Shiyao Wang, Shiwan Zhao, Jiaming Zhou, Yong Qin

Dysarthric speech recognition (DSR) research has witnessed remarkable progress in recent years, evolving from the basic understanding of individual words to the intricate comprehension of sentence-level expressions, all driven by the pressing communication needs of individuals with dysarthria. Nevertheless, the scarcity of available data remains a substantial hurdle, posing a significant challenge to the development of effective sentence-level DSR systems. In response to this issue, dysarthric data augmentation (DDA) has emerged as a highly promising approach. Generative models are frequently employed to generate training data for automatic speech recognition tasks. However, their effectiveness hinges on the ability of the synthesized data to accurately represent the target domain. The wide-ranging variability in pronunciation among dysarthric speakers makes it extremely difficult for models trained on data from existing speakers to produce useful augmented data, especially in zero-shot or one-shot learning settings. To address this limitation, we put forward a novel text-coverage strategy specifically designed for text-matching data synthesis. This innovative strategy allows for efficient zero/one-shot DDA, leading to substantial enhancements in the performance of DSR when dealing with unseen dysarthric speakers. Such improvements are of great significance in practical applications, including dysarthria rehabilitation programs and day-to-day common-sentence communication scenarios.

comment: NCMMSC 2025 oral

☆ SAC: Neural Speech Codec with Semantic-Acoustic Dual-Stream Quantization

Wenxi Chen, Xinsheng Wang, Ruiqi Yan, Yushen Chen, Zhikang Niu, Ziyang Ma, Xiquan Li, Yuzhe Liang, Hanlin Wen, Shunshun Yin, Ming Tao, Xie Chen

Speech codecs that convert continuous speech signals into discrete tokens have become essential for speech language models (SLMs). However, existing codecs struggle to balance high-quality reconstruction with semantically rich representations, limiting their effectiveness in both generative and understanding tasks. In this work, we propose SAC, a neural speech codec with semantic-acoustic dual-stream quantization. By disentangling semantic and acoustic modeling into two dedicated streams, SAC enables each to be optimized for its respective role. Comprehensive evaluations show that SAC achieves strong reconstruction performance across diverse bitrates under both clean and noisy conditions, with particularly high scores on UTMOS and WER, demonstrating superior perceptual quality and intelligibility. Moreover, SAC substantially outperforms state-of-the-art codecs in semantic representation, achieving a level comparable to that of self-supervised learning (SSL) continuous embeddings. Finally, our analysis of speech disentanglement highlights the effectiveness of the dual-stream design, offering new potential for controllable speech applications.

♻ ☆ BINAQUAL: A Full-Reference Objective Localization Similarity Metric for Binaural Audio

Davoud Shariat Panah, Dan Barry, Alessandro Ragano, Jan Skoglund, Andrew Hines

Spatial audio enhances immersion in applications such as virtual reality, augmented reality, gaming, and cinema by creating a three-dimensional auditory experience. Ensuring the spatial fidelity of binaural audio is crucial, given that processes such as compression, encoding, or transmission can alter localization cues. While subjective listening tests like MUSHRA remain the gold standard for evaluating spatial localization quality, they are costly and time-consuming. This paper introduces BINAQUAL, a full-reference objective metric designed to assess localization similarity in binaural audio recordings. BINAQUAL adapts the AMBIQUAL metric, originally developed for localization quality assessment in ambisonics audio format to the binaural domain. We evaluate BINAQUAL across five key research questions, examining its sensitivity to variations in sound source locations, angle interpolations, surround speaker layouts, audio degradations, and content diversity. Results demonstrate that BINAQUAL effectively differentiates between subtle spatial variations and correlates strongly with subjective listening tests, making it a reliable metric for binaural localization quality assessment. The proposed metric provides a robust benchmark for ensuring spatial accuracy in binaural audio processing, paving the way for improved objective evaluations in immersive audio applications.

comment: Accepted for publication in the Journal of Audio Engineering Society (JAES)

☆ Towards Real-Time Generative Speech Restoration with Flow-Matching

Tsun-An Hsieh, Sebastian Braun

Generative models have shown robust performance on speech enhancement and restoration tasks, but most prior approaches operate offline with high latency, making them unsuitable for streaming applications. In this work, we investigate the feasibility of a low-latency, real-time generative speech restoration system based on flow-matching (FM). Our method tackles diverse real-world tasks, including denoising, dereverberation, and generative restoration. The proposed causal architecture without time-downsampling achieves introduces an total latency of only 20 ms, suitable for real-time communication. In addition, we explore a broad set of architectural variations and sampling strategies to ensure effective training and efficient inference. Notably, our flow-matching model maintains high enhancement quality with only 5 number of function evaluations (NFEs) during sampling, achieving similar performance as when using ~20 NFEs under the same conditions. Experimental results indicate that causal FM-based models favor few-step reverse sampling, and smaller backbones degrade with longer reverse trajectories. We further show a side-by-side comparison of FM to typical adversarial-loss-based training for the same model architecture.

☆ Adaptive Deterministic Flow Matching for Target Speaker Extraction

Tsun-An Hsieh, Minje Kim

Generative target speaker extraction (TSE) methods often produce more natural outputs than predictive models. Recent work based on diffusion or flow matching (FM) typically relies on a small, fixed number of reverse steps with a fixed step size. We introduce Adaptive Discriminative Flow Matching TSE (AD-FlowTSE), which extracts the target speech using an adaptive step size. We formulate TSE within the FM paradigm but, unlike prior FM-based speech enhancement and TSE approaches that transport between the mixture (or a normal prior) and the clean-speech distribution, we define the flow between the background and the source, governed by the mixing ratio (MR) of the source and background that creates the mixture. This design enables MR-aware initialization, where the model starts at an adaptive point along the background-source trajectory rather than applying the same reverse schedule across all noise levels. Experiments show that AD-FlowTSE achieves strong TSE with as few as a single step, and that incorporating auxiliary MR estimation further improves target speech accuracy. Together, these results highlight that aligning the transport path with the mixture composition and adapting the step size to noise conditions yields efficient and accurate TSE.

☆ SAKE: Towards Editing Auditory Attribute Knowledge of Large Audio-Language Models

Chih-Kai Yang, Yen-Ting Piao, Tzu-Wen Hsu, Szu-Wei Fu, Zhehuai Chen, Ke-Han Lu, Sung-Feng Huang, Chao-Han Huck Yang, Yu-Chiang Frank Wang, Yun-Nung Chen, Hung-yi Lee

Knowledge editing offers an efficient way to update model knowledge without full retraining, but prior work has concentrated almost exclusively on textual or visual modalities. We introduce SAKE, the first benchmark specifically designed for editing auditory attribute knowledge in Large Audio-Language Models (LALMs). Unlike factual updates, SAKE targets several abstract auditory attributes, capturing knowledge types that go beyond conventional textual and visual domains. We benchmark seven editing methods on two LALMs along four dimensions: reliability, generality, audio/text locality, and portability. Results highlight challenges such as preserving intra-attribute knowledge unrelated to the edit, generalizing edits to multimodal reasoning, and maintaining edits under sequential updates. SAKE provides a principled framework to study how knowledge editing extends to the auditory modalities, opening new directions for maintaining and adapting LALMs in more diverse real-world scenarios.

comment: Work in progress

☆ Investigating Safety Vulnerabilities of Large Audio-Language Models Under Speaker Emotional Variations ICASSP 2026

Bo-Han Feng, Chien-Feng Liu, Yu-Hsuan Li Liang, Chih-Kai Yang, Szu-Wei Fu, Zhehuai Chen, Ke-Han Lu, Sung-Feng Huang, Chao-Han Huck Yang, Yu-Chiang Frank Wang, Yun-Nung Chen, Hung-yi Lee

Large audio-language models (LALMs) extend text-based LLMs with auditory understanding, offering new opportunities for multimodal applications. While their perception, reasoning, and task performance have been widely studied, their safety alignment under paralinguistic variation remains underexplored. This work systematically investigates the role of speaker emotion. We construct a dataset of malicious speech instructions expressed across multiple emotions and intensities, and evaluate several state-of-the-art LALMs. Our results reveal substantial safety inconsistencies: different emotions elicit varying levels of unsafe responses, and the effect of intensity is non-monotonic, with medium expressions often posing the greatest risk. These findings highlight an overlooked vulnerability in LALMs and call for alignment strategies explicitly designed to ensure robustness under emotional variation, a prerequisite for trustworthy deployment in real-world settings.

comment: Submitted to ICASSP 2026

☆ SAC: Neural Speech Codec with Semantic-Acoustic Dual-Stream Quantization

Wenxi Chen, Xinsheng Wang, Ruiqi Yan, Yushen Chen, Zhikang Niu, Ziyang Ma, Xiquan Li, Yuzhe Liang, Hanlin Wen, Shunshun Yin, Ming Tao, Xie Chen

Speech codecs that convert continuous speech signals into discrete tokens have become essential for speech language models (SLMs). However, existing codecs struggle to balance high-quality reconstruction with semantically rich representations, limiting their effectiveness in both generative and understanding tasks. In this work, we propose SAC, a neural speech codec with semantic-acoustic dual-stream quantization. By disentangling semantic and acoustic modeling into two dedicated streams, SAC enables each to be optimized for its respective role. Comprehensive evaluations show that SAC achieves strong reconstruction performance across diverse bitrates under both clean and noisy conditions, with particularly high scores on UTMOS and WER, demonstrating superior perceptual quality and intelligibility. Moreover, SAC substantially outperforms state-of-the-art codecs in semantic representation, achieving a level comparable to that of self-supervised learning (SSL) continuous embeddings. Finally, our analysis of speech disentanglement highlights the effectiveness of the dual-stream design, offering new potential for controllable speech applications.

☆ Schrödinger Bridge Mamba for One-Step Speech Enhancement

Jing Yang, Sirui Wang, Chao Wu, Fan Fan

We propose Schr\"odinger Bridge Mamba (SBM), a new concept of training-inference framework motivated by the inherent compatibility between Schr\"odinger Bridge (SB) training paradigm and selective state-space model Mamba. We exemplify the concept of SBM with an implementation for generative speech enhancement. Experiments on a joint denoising and dereverberation task using four benchmark datasets demonstrate that SBM, with only 1-step inference, outperforms strong baselines with 1-step or iterative inference and achieves the best real-time factor (RTF). Beyond speech enhancement, we discuss the integration of SB paradigm and selective state-space model architecture based on their underlying alignment, which indicates a promising direction for exploring new deep generative models potentially applicable to a broad range of generative tasks. Demo page: https://sbmse.github.io

comment: 5 pages, 1 figure

☆ Audio dequantization using instantaneous frequency

Vojtěch Kovanda, Pavel Rajmic

We present a dequantization method that employs a phase-aware regularizer, originally successfully applied in an audio inpainting problem. The method maintains the temporal continuity of sinusoidal components in the audio signal time-frequency representation and avoids the energy loss artifacts commonly encountered with l1-based regularization approaches. The proposed method is called the Phase-Aware Audio Dequantizer (PHADQ). The methods are evaluated using the objective metric SDR and PEMO-Q ODG.

☆ End-to-end Listen, Look, Speak and Act

Siyin Wang, Wenyi Yu, Xianzhao Chen, Xiaohai Tian, Jun Zhang, Lu Lu, Chao Zhang

Human interaction is inherently multimodal and full-duplex: we listen while watching, speak while acting, and fluidly adapt to turn-taking and interruptions. Realizing these capabilities is essential for building models simulating humans. We present ELLSA (End-to-end Listen, Look, Speak and Act), which, to our knowledge, is the first full-duplex, end-to-end model that simultaneously perceives and generates across vision, text, speech, and action within a single architecture, enabling interaction patterns previously out of reach, yielding more natural, human-like behaviors. At its core is a novel SA-MoE architecture (Self-Attention Mixture-of-Experts) that routes each modality to specialized experts and fuses them through a unified attention backbone. This provides a generalizable solution for joint multimodal perception and concurrent generation, leveraging strong pre-trained components while enabling efficient modality integration and mitigating modality interference. On speech-interaction and robot-manipulation benchmarks, ELLSA matches modality-specific baselines, while uniquely supporting advanced multimodal and full-duplex behaviors such as dialogue and action turn-taking, defective instruction rejection, speaking-while-acting, context-grounded visual question answering, and action barge-ins. We contend that ELLSA represents a step toward more natural and general interactive intelligence, contributing to the broader pursuit of artificial general intelligence. All data, code and model checkpoints will be released upon acceptance.

comment: 22 pages, 8 figures

♻ ☆ BINAQUAL: A Full-Reference Objective Localization Similarity Metric for Binaural Audio

Davoud Shariat Panah, Dan Barry, Alessandro Ragano, Jan Skoglund, Andrew Hines

Spatial audio enhances immersion in applications such as virtual reality, augmented reality, gaming, and cinema by creating a three-dimensional auditory experience. Ensuring the spatial fidelity of binaural audio is crucial, given that processes such as compression, encoding, or transmission can alter localization cues. While subjective listening tests like MUSHRA remain the gold standard for evaluating spatial localization quality, they are costly and time-consuming. This paper introduces BINAQUAL, a full-reference objective metric designed to assess localization similarity in binaural audio recordings. BINAQUAL adapts the AMBIQUAL metric, originally developed for localization quality assessment in ambisonics audio format to the binaural domain. We evaluate BINAQUAL across five key research questions, examining its sensitivity to variations in sound source locations, angle interpolations, surround speaker layouts, audio degradations, and content diversity. Results demonstrate that BINAQUAL effectively differentiates between subtle spatial variations and correlates strongly with subjective listening tests, making it a reliable metric for binaural localization quality assessment. The proposed metric provides a robust benchmark for ensuring spatial accuracy in binaural audio processing, paving the way for improved objective evaluations in immersive audio applications.

comment: Accepted for publication in the Journal of Audio Engineering Society (JAES)

♻ ☆ SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement NeurIPS2025

Chenyu Yang, Shuai Wang, Hangting Chen, Wei Tan, Jianwei Yu, Haizhou Li

Generating music with coherent structure, harmonious instrumental and vocal elements remains a significant challenge in song generation. Existing language models and diffusion-based methods often struggle to balance global coherence with local fidelity, resulting in outputs that lack musicality or suffer from incoherent progression and mismatched lyrics. This paper introduces $\textbf{SongBloom}$, a novel framework for full-length song generation that leverages an interleaved paradigm of autoregressive sketching and diffusion-based refinement. SongBloom employs an autoregressive diffusion model that combines the high fidelity of diffusion models with the scalability of language models. Specifically, it gradually extends a musical sketch from short to long and refines the details from coarse to fine-grained. The interleaved generation paradigm effectively integrates prior semantic and acoustic context to guide the generation process. Experimental results demonstrate that SongBloom outperforms existing methods across both subjective and objective metrics and achieves performance comparable to the state-of-the-art commercial music generation platforms. Audio samples are available on our demo page: https://cypress-yang.github.io/SongBloom_demo. The code and model weights have been released on https://github.com/Cypress-Yang/SongBloom .

comment: Accepted by NeurIPS2025

♻ ☆ A Self-Attention-Driven Deep Denoiser Model for Real Time Lung Sound Denoising in Noisy Environments

Samiul Based Shuvo, Syed Samiul Alam, Taufiq Hasan

Objective: Lung auscultation is a valuable tool in diagnosing and monitoring various respiratory diseases. However, lung sounds (LS) are significantly affected by numerous sources of contamination, especially when recorded in real-world clinical settings. Conventional denoising models prove impractical for LS denoising, primarily owing to spectral overlap complexities arising from diverse noise sources. To address this issue, we propose a specialized deep-learning model (Uformer) for lung sound denoising. Methods: The proposed Uformer model is constituted of three modules: a Convolutional Neural Network (CNN) encoder module, dedicated to extracting latent features; a Transformer encoder module, employed to further enhance the encoding of unique LS features and effectively capture intricate long-range dependencies; and a CNN decoder module, employed to generate the denoised signals. An ablation study was performed in order to find the most optimal architecture. Results: The performance of the proposed Uformer model was evaluated on lung sounds induced with different types of synthetic and real-world noises. Lung sound signals of -12 dB to 15 dB signal-to-noise ratio (SNR) were considered in testing experiments. The proposed model showed an average SNR improvement of 16.51 dB when evaluated with -12 dB LS signals. Our end-to-end model, with an average SNR improvement of 19.31 dB, outperforms the existing model when evaluated with ambient noise and fewer parameters. Conclusion: Based on the qualitative and quantitative findings in this study, it can be stated that Uformer is robust and generalized to be used in assisting the monitoring of respiratory conditions.

♻ ☆ Guitar Tone Morphing by Diffusion-based Model SC 2025

Kuan-Yu Chen, Kuan-Lin Chen, Yu-Chieh Yu, Jian-Jiun Ding

In Music Information Retrieval (MIR), modeling and transforming the tone of musical instruments, particularly electric guitars, has gained increasing attention due to the richness of the instrument tone and the flexibility of expression. Tone morphing enables smooth transitions between different guitar sounds, giving musicians greater freedom to explore new textures and personalize their performances. This study explores learning-based approaches for guitar tone morphing, beginning with LoRA fine-tuning to improve the model performance on limited data. Moreover, we introduce a simpler method, named spherical interpolation using Music2Latent. It yields significantly better results than the more complex fine-tuning approach. Experiments show that the proposed architecture generates smoother and more natural tone transitions, making it a practical and efficient tool for music production and real-time audio effects.

comment: 5 pages, accepted to the APSIPA ASC 2025

☆ Hallucination Benchmark for Speech Foundation Models

Alkis Koudounas, Moreno La Quatra, Manuel Giollo, Sabato Marco Siniscalchi, Elena Baralis

Hallucinations in automatic speech recognition (ASR) systems refer to fluent and coherent transcriptions produced by neural ASR models that are completely unrelated to the underlying acoustic input (i.e., the speech signal). While similar to conventional decoding errors in potentially compromising the usability of transcriptions for downstream applications, hallucinations can be more detrimental due to their preservation of syntactically and semantically plausible structure. This apparent coherence can mislead subsequent processing stages and introduce serious risks, particularly in critical domains such as healthcare and law. Conventional evaluation metrics are primarily centered on error-based metrics and fail to distinguish between phonetic inaccuracies and hallucinations. Consequently, there is a critical need for new evaluation frameworks that can effectively identify and assess models with a heightened propensity for generating hallucinated content. To this end, we introduce SHALLOW, the first benchmark framework that systematically categorizes and quantifies hallucination phenomena in ASR along four complementary axes: lexical, phonetic, morphological, and semantic. We define targeted metrics within each category to produce interpretable profiles of model behavior. Through evaluation across various architectures and speech domains, we have found that SHALLOW metrics correlate strongly with word error rate (WER) when recognition quality is high (i.e., low WER). Still, this correlation weakens substantially as WER increases. SHALLOW, therefore, captures fine-grained error patterns that WER fails to distinguish under degraded and challenging conditions. Our framework supports specific diagnosis of model weaknesses and provides feedback for model improvement beyond what aggregate error rates can offer.

comment: Under Review

☆ Interpreting the Dimensions of Speaker Embedding Space

Mark Huckvale

Speaker embeddings are widely used in speaker verification systems and other applications where it is useful to characterise the voice of a speaker with a fixed-length vector. These embeddings tend to be treated as "black box" encodings, and how they relate to conventional acoustic and phonetic dimensions of voices has not been widely studied. In this paper we investigate how state-of-the-art speaker embedding systems represent the acoustic characteristics of speakers as described by conventional acoustic descriptors, age, and gender. Using a large corpus of 10,000 speakers and three embedding systems we show that a small set of 9 acoustic parameters chosen to be "interpretable" predict embeddings about the same as 7 principal components, corresponding to over 50% of variance in the data. We show that some principal dimensions operate differently for male and female speakers, suggesting there is implicit gender recognition within the embedding systems. However we show that speaker age is not well captured by embeddings, suggesting opportunities exist for improvements in their calculation.

☆ Probing the Hidden Talent of ASR Foundation Models for L2 English Oral Assessment

Fu-An Chao, Bi-Cheng Yan, Berlin Chen

In this paper, we explore the untapped potential of Whisper, a well-established automatic speech recognition (ASR) foundation model, in the context of L2 spoken language assessment (SLA). Unlike prior studies that extrinsically analyze transcriptions produced by Whisper, our approach goes a step further to probe its latent capabilities by extracting acoustic and linguistic features from hidden representations. With only a lightweight classifier being trained on top of Whisper's intermediate and final outputs, our method achieves strong performance on the GEPT picture-description dataset, outperforming existing cutting-edge baselines, including a multimodal approach. Furthermore, by incorporating image and text-prompt information as auxiliary relevance cues, we demonstrate additional performance gains. Finally, we conduct an in-depth analysis of Whisper's embeddings, which reveals that, even without task-specific fine-tuning, the model intrinsically encodes both ordinal proficiency patterns and semantic aspects of speech, highlighting its potential as a powerful foundation for SLA and other spoken language understanding tasks.

☆ Transmission of High-Amplitude Sound through Leakages of Ill-fitting Earplugs

Haocheng Yu, Krishan K. Ahuja, Lakshmi N. Sankar, Spencer H. Bryngelson

High sound pressure levels (SPL) pose notable risks in loud environments, particularly due to noise-induced hearing loss. Ill-fitting earplugs often lead to sound leakage, a phenomenon this study seeks to investigate. To validate our methodology, we first obtained computational and experimental acoustic transmission data for stand-alone slit resonators and orifices, for which extensive published data are readily available for comparison. We then examined the frequency-dependent acoustic power absorption coefficient and transmission loss (TL) across various leakage geometries, modeled using different orifice diameters. Experimental approaches spanned a frequency range of 1--5 kHz under SPL conditions of 120--150 dB. Key findings reveal that unsealed silicone rubber earplugs demonstrate an average TL reduction of approximately 18 dB at an overall incident SPL (OISPL) of 120 dB. Direct numerical simulations further highlight SPL-dependent acoustic dissipation mechanisms, showing the conversion of acoustic energy into vorticity in ill-fitting earplug models at an OISPL of 150 dB. These results highlight the role of earplug design for high-sound-pressure-level environments.

☆ MuseTok: Symbolic Music Tokenization for Generation and Semantic Understanding

Jingyue Huang, Zachary Novack, Phillip Long, Yupeng Hou, Ke Chen, Taylor Berg-Kirkpatrick, Julian McAuley

Discrete representation learning has shown promising results across various domains, including generation and understanding in image, speech and language. Inspired by these advances, we propose MuseTok, a tokenization method for symbolic music, and investigate its effectiveness in both music generation and understanding tasks. MuseTok employs the residual vector quantized-variational autoencoder (RQ-VAE) on bar-wise music segments within a Transformer-based encoder-decoder framework, producing music codes that achieve high-fidelity music reconstruction and accurate understanding of music theory. For comprehensive evaluation, we apply MuseTok to music generation and semantic understanding tasks, including melody extraction, chord recognition, and emotion recognition. Models incorporating MuseTok outperform previous representation learning baselines in semantic understanding while maintaining comparable performance in content generation. Furthermore, qualitative analyses on MuseTok codes, using ground-truth categories and synthetic datasets, reveal that MuseTok effectively captures underlying musical concepts from large music collections.

♻ ☆ CoVoMix2: Advancing Zero-Shot Dialogue Generation with Fully Non-Autoregressive Flow Matching

Leying Zhang, Yao Qian, Xiaofei Wang, Manthan Thakker, Dongmei Wang, Jianwei Yu, Haibin Wu, Yuxuan Hu, Jinyu Li, Yanmin Qian, Sheng Zhao

Generating natural-sounding, multi-speaker dialogue is crucial for applications such as podcast creation, virtual agents, and multimedia content generation. However, existing systems struggle to maintain speaker consistency, model overlapping speech, and synthesize coherent conversations efficiently. In this paper, we introduce CoVoMix2, a fully non-autoregressive framework for zero-shot multi-talker dialogue generation. CoVoMix2 directly predicts mel-spectrograms from multi-stream transcriptions using a flow-matching-based generative model, eliminating the reliance on intermediate token representations. To better capture realistic conversational dynamics, we propose transcription-level speaker disentanglement, sentence-level alignment, and prompt-level random masking strategies. Our approach achieves state-of-the-art performance, outperforming strong baselines like MoonCast and Sesame in speech quality, speaker consistency, and inference speed. Notably, CoVoMix2 operates without requiring transcriptions for the prompt and supports controllable dialogue generation, including overlapping speech and precise timing control, demonstrating strong generalizability to real-world speech generation scenarios.

comment: Neural Information Processing Systems 2025, poster

♻ ☆ VGGSounder: Audio-Visual Evaluations for Foundation Models ICCV

Daniil Zverev, Thaddäus Wiedemer, Ameya Prabhu, Matthias Bethge, Wieland Brendel, A. Sophia Koepke

The emergence of audio-visual foundation models underscores the importance of reliably assessing their multi-modal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric.

comment: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) 2025

☆ Hallucination Benchmark for Speech Foundation Models

Alkis Koudounas, Moreno La Quatra, Manuel Giollo, Sabato Marco Siniscalchi, Elena Baralis

Hallucinations in automatic speech recognition (ASR) systems refer to fluent and coherent transcriptions produced by neural ASR models that are completely unrelated to the underlying acoustic input (i.e., the speech signal). While similar to conventional decoding errors in potentially compromising the usability of transcriptions for downstream applications, hallucinations can be more detrimental due to their preservation of syntactically and semantically plausible structure. This apparent coherence can mislead subsequent processing stages and introduce serious risks, particularly in critical domains such as healthcare and law. Conventional evaluation metrics are primarily centered on error-based metrics and fail to distinguish between phonetic inaccuracies and hallucinations. Consequently, there is a critical need for new evaluation frameworks that can effectively identify and assess models with a heightened propensity for generating hallucinated content. To this end, we introduce SHALLOW, the first benchmark framework that systematically categorizes and quantifies hallucination phenomena in ASR along four complementary axes: lexical, phonetic, morphological, and semantic. We define targeted metrics within each category to produce interpretable profiles of model behavior. Through evaluation across various architectures and speech domains, we have found that SHALLOW metrics correlate strongly with word error rate (WER) when recognition quality is high (i.e., low WER). Still, this correlation weakens substantially as WER increases. SHALLOW, therefore, captures fine-grained error patterns that WER fails to distinguish under degraded and challenging conditions. Our framework supports specific diagnosis of model weaknesses and provides feedback for model improvement beyond what aggregate error rates can offer.

comment: Under Review

☆ Interpreting the Dimensions of Speaker Embedding Space

Mark Huckvale

Speaker embeddings are widely used in speaker verification systems and other applications where it is useful to characterise the voice of a speaker with a fixed-length vector. These embeddings tend to be treated as "black box" encodings, and how they relate to conventional acoustic and phonetic dimensions of voices has not been widely studied. In this paper we investigate how state-of-the-art speaker embedding systems represent the acoustic characteristics of speakers as described by conventional acoustic descriptors, age, and gender. Using a large corpus of 10,000 speakers and three embedding systems we show that a small set of 9 acoustic parameters chosen to be "interpretable" predict embeddings about the same as 7 principal components, corresponding to over 50% of variance in the data. We show that some principal dimensions operate differently for male and female speakers, suggesting there is implicit gender recognition within the embedding systems. However we show that speaker age is not well captured by embeddings, suggesting opportunities exist for improvements in their calculation.

☆ Audio-Visual Speech Enhancement for Spatial Audio - Spatial-VisualVoice and the MAVE Database

Danielle Yaffe, Ferdinand Campe, Prachi Sharma, Dorothea Kolossa, Boaz Rafaely

Audio-visual speech enhancement (AVSE) has been found to be particularly useful at low signal-to-noise (SNR) ratios due to the immunity of the visual features to acoustic noise. However, a significant gap exists in AVSE methods tailored to enhance spatial audio under low-SNR conditions. The latter is of growing interest with augmented reality applications. To address this gap, we present a multi-channel AVSE framework based on VisualVoice that leverages spatial cues from microphone arrays and visual information for enhancing the target speaker in noisy environments. We also introduce MAVe, a novel database containing multi-channel audio-visual signals in controlled, reproducible room conditions across a wide range of SNR levels. Experiments demonstrate that the proposed method consistently achieves significant gains in SI-SDR, STOI, and PESQ, particularly in low SNRs. Binaural signal analysis further confirms the preservation of spatial cues and intelligibility.

☆ Probing the Hidden Talent of ASR Foundation Models for L2 English Oral Assessment

Fu-An Chao, Bi-Cheng Yan, Berlin Chen

In this paper, we explore the untapped potential of Whisper, a well-established automatic speech recognition (ASR) foundation model, in the context of L2 spoken language assessment (SLA). Unlike prior studies that extrinsically analyze transcriptions produced by Whisper, our approach goes a step further to probe its latent capabilities by extracting acoustic and linguistic features from hidden representations. With only a lightweight classifier being trained on top of Whisper's intermediate and final outputs, our method achieves strong performance on the GEPT picture-description dataset, outperforming existing cutting-edge baselines, including a multimodal approach. Furthermore, by incorporating image and text-prompt information as auxiliary relevance cues, we demonstrate additional performance gains. Finally, we conduct an in-depth analysis of Whisper's embeddings, which reveals that, even without task-specific fine-tuning, the model intrinsically encodes both ordinal proficiency patterns and semantic aspects of speech, highlighting its potential as a powerful foundation for SLA and other spoken language understanding tasks.

☆ Transmission of High-Amplitude Sound through Leakages of Ill-fitting Earplugs

Haocheng Yu, Krishan K. Ahuja, Lakshmi N. Sankar, Spencer H. Bryngelson

High sound pressure levels (SPL) pose notable risks in loud environments, particularly due to noise-induced hearing loss. Ill-fitting earplugs often lead to sound leakage, a phenomenon this study seeks to investigate. To validate our methodology, we first obtained computational and experimental acoustic transmission data for stand-alone slit resonators and orifices, for which extensive published data are readily available for comparison. We then examined the frequency-dependent acoustic power absorption coefficient and transmission loss (TL) across various leakage geometries, modeled using different orifice diameters. Experimental approaches spanned a frequency range of 1--5 kHz under SPL conditions of 120--150 dB. Key findings reveal that unsealed silicone rubber earplugs demonstrate an average TL reduction of approximately 18 dB at an overall incident SPL (OISPL) of 120 dB. Direct numerical simulations further highlight SPL-dependent acoustic dissipation mechanisms, showing the conversion of acoustic energy into vorticity in ill-fitting earplug models at an OISPL of 150 dB. These results highlight the role of earplug design for high-sound-pressure-level environments.

☆ MuseTok: Symbolic Music Tokenization for Generation and Semantic Understanding

Jingyue Huang, Zachary Novack, Phillip Long, Yupeng Hou, Ke Chen, Taylor Berg-Kirkpatrick, Julian McAuley

Discrete representation learning has shown promising results across various domains, including generation and understanding in image, speech and language. Inspired by these advances, we propose MuseTok, a tokenization method for symbolic music, and investigate its effectiveness in both music generation and understanding tasks. MuseTok employs the residual vector quantized-variational autoencoder (RQ-VAE) on bar-wise music segments within a Transformer-based encoder-decoder framework, producing music codes that achieve high-fidelity music reconstruction and accurate understanding of music theory. For comprehensive evaluation, we apply MuseTok to music generation and semantic understanding tasks, including melody extraction, chord recognition, and emotion recognition. Models incorporating MuseTok outperform previous representation learning baselines in semantic understanding while maintaining comparable performance in content generation. Furthermore, qualitative analyses on MuseTok codes, using ground-truth categories and synthetic datasets, reveal that MuseTok effectively captures underlying musical concepts from large music collections.

♻ ☆ CoVoMix2: Advancing Zero-Shot Dialogue Generation with Fully Non-Autoregressive Flow Matching

Leying Zhang, Yao Qian, Xiaofei Wang, Manthan Thakker, Dongmei Wang, Jianwei Yu, Haibin Wu, Yuxuan Hu, Jinyu Li, Yanmin Qian, Sheng Zhao

Generating natural-sounding, multi-speaker dialogue is crucial for applications such as podcast creation, virtual agents, and multimedia content generation. However, existing systems struggle to maintain speaker consistency, model overlapping speech, and synthesize coherent conversations efficiently. In this paper, we introduce CoVoMix2, a fully non-autoregressive framework for zero-shot multi-talker dialogue generation. CoVoMix2 directly predicts mel-spectrograms from multi-stream transcriptions using a flow-matching-based generative model, eliminating the reliance on intermediate token representations. To better capture realistic conversational dynamics, we propose transcription-level speaker disentanglement, sentence-level alignment, and prompt-level random masking strategies. Our approach achieves state-of-the-art performance, outperforming strong baselines like MoonCast and Sesame in speech quality, speaker consistency, and inference speed. Notably, CoVoMix2 operates without requiring transcriptions for the prompt and supports controllable dialogue generation, including overlapping speech and precise timing control, demonstrating strong generalizability to real-world speech generation scenarios.

comment: Neural Information Processing Systems 2025, poster

♻ ☆ VGGSounder: Audio-Visual Evaluations for Foundation Models ICCV

Daniil Zverev, Thaddäus Wiedemer, Ameya Prabhu, Matthias Bethge, Wieland Brendel, A. Sophia Koepke

The emergence of audio-visual foundation models underscores the importance of reliably assessing their multi-modal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric.

comment: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) 2025

♻ ☆ SHANKS: Simultaneous Hearing and Thinking for Spoken Language Models

Cheng-Han Chiang, Xiaofei Wang, Linjie Li, Chung-Ching Lin, Kevin Lin, Shujie Liu, Zhendong Wang, Zhengyuan Yang, Hung-yi Lee, Lijuan Wang

Current large language models (LLMs) and spoken language models (SLMs) begin thinking and taking actions only after the user has finished their turn. This prevents the model from interacting during the user's turn and can lead to high response latency while it waits to think. Consequently, thinking after receiving the full input is not suitable for speech-to-speech interaction, where real-time, low-latency exchange is important. We address this by noting that humans naturally "think while listening." In this paper, we propose SHANKS, a general inference framework that enables SLMs to generate unspoken chain-of-thought reasoning while listening to the user input. SHANKS streams the input speech in fixed-duration chunks and, as soon as a chunk is received, generates unspoken reasoning based on all previous speech and reasoning, while the user continues speaking. SHANKS uses this unspoken reasoning to decide whether to interrupt the user and to make tool calls to complete the task. We demonstrate that SHANKS enhances real-time user-SLM interaction in two scenarios: (1) when the user is presenting a step-by-step solution to a math problem, SHANKS can listen, reason, and interrupt when the user makes a mistake, achieving 37.1% higher interruption accuracy than a baseline that interrupts without thinking; and (2) in a tool-augmented dialogue, SHANKS can complete 56.9% of the tool calls before the user finishes their turn. Overall, SHANKS moves toward models that keep thinking throughout the conversation, not only after a turn ends. Animated illustrations of Shanks can be found at https://d223302.github.io/SHANKS/

comment: Work in progress

☆ Sound Clouds: Exploring ambient intelligence in public spaces to elicit deep human experience of awe, wonder, and beauty NeurIPS

Chengzhi Zhang, Dashiel Carrera, Daksh Kapoor, Jasmine Kaur, Jisu Kim, Brian Magerko

While the ambient intelligence (AmI) systems we encounter in our daily lives, including security monitoring and energy-saving systems, typically serve pragmatic purposes, we wonder how we can design and implement ambient artificial intelligence experiences in public spaces that elicit deep human feelings of awe, wonder, and beauty. As a manifestation, we introduce Sound Clouds, an immersive art installation that generates live music based on participants' interaction with several human-height spheres. Our installation serves as a provocation into future ambient intelligence that provokes, not limits, the future possibilities of AmI.

comment: 4 pages, Artwork accepted by NeurIPS Creative AI Track 2025

☆ SpikeVox: Towards Energy-Efficient Speech Therapy Framework with Spike-driven Generative Language Models

Rachmad Vidya Wicaksana Putra, Aadithyan Rajesh Nair, Muhammad Shafique

Speech disorders can significantly affect the patients capability to communicate, learn, and socialize. However, existing speech therapy solutions (e.g., therapist or tools) are still limited and costly, hence such solutions remain inadequate for serving millions of patients worldwide. To address this, state-of-the-art methods employ neural network (NN) algorithms to help accurately detecting speech disorders. However, these methods do not provide therapy recommendation as feedback, hence providing partial solution for patients. Moreover, these methods incur high energy consumption due to their complex and resource-intensive NN processing, hence hindering their deployments on low-power/energy platforms (e.g., smartphones). Toward this, we propose SpikeVox, a novel framework for enabling energy-efficient speech therapy solutions through spike-driven generative language model. Specifically, SpikeVox employs a speech recognition module to perform highly accurate speech-to-text conversion; leverages a spike-driven generative language model to efficiently perform pattern analysis for speech disorder detection and generates suitable exercises for therapy; provides guidance on correct pronunciation as feedback; as well as utilizes the REST API to enable seamless interaction for users. Experimental results demonstrate that SpikeVox achieves 88% confidence level on average in speech disorder recognition, while providing a complete feedback for therapy exercises. Therefore, SpikeVox provides a comprehensive framework for energy-efficient speech therapy solutions, and potentially addresses the significant global speech therapy access gap.

comment: Accepted at the IEEE Biomedical Circuits and Systems Conference (BioCAS) 2025, Abu Dhabi, UAE

☆ Quantization-Based Score Calibration for Few-Shot Keyword Spotting with Dynamic Time Warping in Noisy Environments

Kevin Wilkinghoff, Alessia Cornaggia-Urrigshardt, Zheng-Hua Tan

Detecting occurrences of keywords with keyword spotting (KWS) systems requires thresholding continuous detection scores. Selecting appropriate thresholds is a non-trivial task, typically relying on optimizing the performance on a validation dataset. However, such greedy threshold selection often leads to suboptimal performance on unseen data, particularly in varying or noisy acoustic environments or few-shot settings. In this work, we investigate detection threshold estimation for template-based open-set few-shot KWS using dynamic time warping on noisy speech data. To mitigate the performance degradation caused by suboptimal thresholds, we propose a score calibration approach consisting of two different steps: quantizing embeddings and normalizing detection scores using the quantization error prior to thresholding. Experiments on KWS-DailyTalk with simulated high frequency radio channels show that the proposed calibration approach simplifies the choice of detection thresholds and significantly improves the resulting performance.

☆ DroneAudioset: An Audio Dataset for Drone-based Search and Rescue

Chitralekha Gupta, Soundarya Ramesh, Praveen Sasikumar, Kian Peen Yeo, Suranga Nanayakkara

Unmanned Aerial Vehicles (UAVs) or drones, are increasingly used in search and rescue missions to detect human presence. Existing systems primarily leverage vision-based methods which are prone to fail under low-visibility or occlusion. Drone-based audio perception offers promise but suffers from extreme ego-noise that masks sounds indicating human presence. Existing datasets are either limited in diversity or synthetic, lacking real acoustic interactions, and there are no standardized setups for drone audition. To this end, we present DroneAudioset (The dataset is publicly available at https://huggingface.co/datasets/ahlab-drone-project/DroneAudioSet/ under the MIT license), a comprehensive drone audition dataset featuring 23.5 hours of annotated recordings, covering a wide range of signal-to-noise ratios (SNRs) from -57.2 dB to -2.5 dB, across various drone types, throttles, microphone configurations as well as environments. The dataset enables development and systematic evaluation of noise suppression and classification methods for human-presence detection under challenging conditions, while also informing practical design considerations for drone audition systems, such as microphone placement trade-offs, and development of drone noise-aware audio processing. This dataset is an important step towards enabling design and deployment of drone-audition systems.

comment: Accepted in Neurips (Datasets and Benchmarks Track) 2025. The first two authors are equal contributors

☆ Extending Audio Context for Long-Form Understanding in Large Audio-Language Models

Yuatyong Chaichana, Pittawat Taveekitworachai, Warit Sirichotedumrong, Potsawee Manakul, Kunat Pipatanakul

Large Audio-Language Models (LALMs) are often constrained by short audio context windows, even when their text backbones support long contexts, limiting long-form audio understanding. Prior work has introduced context-extension methods (e.g. YaRN) on unimodal LLMs, yet their application to LALMs remains unexplored. First, building on RoPE-based context extension, we introduce Partial YaRN, a training-free, audio-only extension method that modifies only audio token positions, leaving text positions intact to preserve the base LLM's text capabilities. Second, we propose Virtual Longform Audio Training (VLAT), a training strategy that extends Partial YaRN into a training-time positional augmentation. VLAT simulates diverse audio lengths during training, enabling generalization to inputs far longer than those seen in training and improving robustness for long-context audio understanding. Our experiments on SALMONN and Qwen2-Audio show that Partial YaRN outperforms the original models across wide range of settings, and VLAT training strategy provides substantial improvement, achieving strong performance on long audio of unseen lengths.

☆ LongCat-Audio-Codec: An Audio Tokenizer and Detokenizer Solution Designed for Speech Large Language Models

Xiaohan Zhao, Hongyu Xiang, Shengze Ye, Song Li, Zhengkun Tian, Guanyu Chen, Ke Ding, Guanglu Wan

This paper presents LongCat-Audio-Codec, an audio tokenizer and detokenizer solution designed for industrial grade end-to-end speech large language models. By leveraging a decoupled model architecture and a multistage training strategy, LongCat-Audio-Codec exhibits robust semantic modeling capabilities, flexible acoustic feature extraction capabilities, and low-latency streaming synthesis capabilities. It encodes speech at an ultra-low frame rate of 16.67 Hz, with a minimum bitrate of 0.43 kbps and a maximum bitrate of 0.87 kbps. Evaluation results demonstrate that LongCat-Audio-Codec achieves strong speech intelligibility and is capable of synthesizing highquality speech at low bitrate, thus effectively balancing coding efficiency and decoding quality. The inference code and model checkpoints of LongCat-Audio-Codec are available at: https://github.com/meituan-longcat/LongCat-Audio-Codec.

♻ ☆ Summarizing Speech: A Comprehensive Survey EMNLP 2025

Fabian Retkowski, Maike Züfle, Andreas Sudmann, Dinah Pfau, Shinji Watanabe, Jan Niehues, Alexander Waibel

Speech summarization has become an essential tool for efficiently managing and accessing the growing volume of spoken and audiovisual content. However, despite its increasing importance, speech summarization remains loosely defined. The field intersects with several research areas, including speech recognition, text summarization, and specific applications like meeting summarization. This survey not only examines existing datasets and evaluation protocols, which are crucial for assessing the quality of summarization approaches, but also synthesizes recent developments in the field, highlighting the shift from traditional systems to advanced models like fine-tuned cascaded architectures and end-to-end solutions. In doing so, we surface the ongoing challenges, such as the need for realistic evaluation benchmarks, multilingual datasets, and long-context handling.

comment: Accepted to EMNLP 2025

♻ ☆ MRSAudio: A Large-Scale Multimodal Recorded Spatial Audio Dataset with Refined Annotations

Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Xintong Hu, Yu Zhang, Li Tang, Rui Yang, Han Wang, Zongbao Zhang, Yuhan Wang, Yixuan Chen, Hankun Xu, Ke Xu, Pengfei Fan, Zhetao Chen, Yanhao Yu, Qiange Huang, Fei Wu, Zhou Zhao

Humans rely on multisensory integration to perceive spatial environments, where auditory cues enable sound source localization in three-dimensional space. Despite the critical role of spatial audio in immersive technologies such as VR/AR, most existing multimodal datasets provide only monaural audio, which limits the development of spatial audio generation and understanding. To address these challenges, we introduce MRSAudio, a large-scale multimodal spatial audio dataset designed to advance research in spatial audio understanding and generation. MRSAudio spans four distinct components: MRSLife, MRSSpeech, MRSMusic, and MRSSing, covering diverse real-world scenarios. The dataset includes synchronized binaural and ambisonic audio, exocentric and egocentric video, motion trajectories, and fine-grained annotations such as transcripts, phoneme boundaries, lyrics, scores, and prompts. To demonstrate the utility and versatility of MRSAudio, we establish five foundational tasks: audio spatialization, and spatial text to speech, spatial singing voice synthesis, spatial music generation and sound event localization and detection. Results show that MRSAudio enables high-quality spatial modeling and supports a broad range of spatial audio research. Demos and dataset access are available at https://mrsaudio.github.io.

comment: 24 pages

♻ ☆ Benchmarking Fake Voice Detection in the Fake Voice Generation Arms Race

Xutao Mao, Ke Li, Cameron Baird, Ezra Xuanru Tao, Dan Lin

The rapid advancement of fake voice generation technology has ignited a race with detection systems, creating an urgent need to secure the audio ecosystem. However, existing benchmarks suffer from a critical limitation: they typically aggregate diverse fake voice samples into a single dataset for evaluation. This practice masks method-specific artifacts and obscures the varying performance of detectors against different generation paradigms, preventing a nuanced understanding of their true vulnerabilities. To address this gap, we introduce the first ecosystem-level benchmark that systematically evaluates the interplay between 17 state-of-the-art fake voice generators and 8 leading detectors through a novel one-to-one evaluation protocol. This fine-grained analysis exposes previously hidden vulnerabilities and sensitivities that are missed by traditional aggregated testing. We also propose unified scoring systems to quantify both the evasiveness of generators and the robustness of detectors, enabling fair and direct comparisons. Our extensive cross-domain evaluation reveals that modern generators, particularly those based on neural audio codecs and flow matching, consistently evade top-tier detectors. We found that no single detector is universally robust; their effectiveness varies dramatically depending on the generator's architecture, highlighting a significant generalization gap in current defenses. This work provides a more realistic assessment of the threat landscape and offers actionable insights for building the next generation of detection systems.

♻ ☆ BandCondiNet: Parallel Transformers-based Conditional Popular Music Generation with Multi-View Features

Jing Luo, Xinyu Yang, Dorien Herremans

Conditional music generation offers significant advantages in terms of user convenience and control, presenting great potential in AI-generated content research. However, building conditional generative systems for multitrack popular songs presents three primary challenges: insufficient fidelity of input conditions, poor structural modeling, and inadequate inter-track harmony learning in generative models. To address these issues, we propose BandCondiNet, a conditional model based on parallel Transformers, designed to process the multiple music sequences and generate high-quality multitrack samples. Specifically, we propose multi-view features across time and instruments as high-fidelity conditions. Moreover, we propose two specialized modules for BandCondiNet: Structure Enhanced Attention (SEA) to strengthen the musical structure, and Cross-Track Transformer (CTT) to enhance inter-track harmony. We conducted both objective and subjective evaluations on two popular music datasets with different sequence lengths. Objective results on the shorter dataset show that BandCondiNet outperforms other conditional models in 9 out of 10 metrics related to fidelity and inference speed, with the exception of Chord Accuracy. On the longer dataset, BandCondiNet surpasses all conditional models across all 10 metrics. Subjective evaluations across four criteria reveal that BandCondiNet trained on the shorter dataset performs best in Richness and performs comparably to state-of-the-art models in the other three criteria, while significantly outperforming them across all criteria when trained on the longer dataset. To further expand the application scope of BandCondiNet, future work should focus on developing an advanced conditional model capable of adapting to more user-friendly input conditions and supporting flexible instrumentation.

comment: To appear in ESWA. Demo page: https://chinglohsiu.github.io/files/bandcondinet.html

♻ ☆ EmoSphere-SER: Enhancing Speech Emotion Recognition Through Spherical Representation with Auxiliary Classification

Deok-Hyeon Cho, Hyung-Seok Oh, Seung-Bin Kim, Seong-Whan Lee

Speech emotion recognition predicts a speaker's emotional state from speech signals using discrete labels or continuous dimensions such as arousal, valence, and dominance (VAD). We propose EmoSphere-SER, a joint model that integrates spherical VAD region classification to guide VAD regression for improved emotion prediction. In our framework, VAD values are transformed into spherical coordinates that are divided into multiple spherical regions, and an auxiliary classification task predicts which spherical region each point belongs to, guiding the regression process. Additionally, we incorporate a dynamic weighting scheme and a style pooling layer with multi-head self-attention to capture spectral and temporal dynamics, further boosting performance. This combined training strategy reinforces structured learning and improves prediction consistency. Experimental results show that our approach exceeds baseline methods, confirming the validity of the proposed framework.

comment: Proceedings of Interspeech 2025

♻ ☆ DiEmo-TTS: Disentangled Emotion Representations via Self-Supervised Distillation for Cross-Speaker Emotion Transfer in Text-to-Speech

Deok-Hyeon Cho, Hyung-Seok Oh, Seung-Bin Kim, Seong-Whan Lee

Cross-speaker emotion transfer in speech synthesis relies on extracting speaker-independent emotion embeddings for accurate emotion modeling without retaining speaker traits. However, existing timbre compression methods fail to fully separate speaker and emotion characteristics, causing speaker leakage and degraded synthesis quality. To address this, we propose DiEmo-TTS, a self-supervised distillation method to minimize emotional information loss and preserve speaker identity. We introduce cluster-driven sampling and information perturbation to preserve emotion while removing irrelevant factors. To facilitate this process, we propose an emotion clustering and matching approach using emotional attribute prediction and speaker embeddings, enabling generalization to unlabeled data. Additionally, we designed a dual conditioning transformer to integrate style features better. Experimental results confirm the effectiveness of our method in learning speaker-irrelevant emotion embeddings.

comment: Proceedings of Interspeech 2025

♻ ☆ Improving Inference-Time Optimisation for Vocal Effects Style Transfer with a Gaussian Prior SP

Chin-Yun Yu, Marco A. Martínez-Ramírez, Junghyun Koo, Wei-Hsiang Liao, Yuki Mitsufuji, György Fazekas

Style Transfer with Inference-Time Optimisation (ST-ITO) is a recent approach for transferring the applied effects of a reference audio to an audio track. It optimises the effect parameters to minimise the distance between the style embeddings of the processed audio and the reference. However, this method treats all possible configurations equally and relies solely on the embedding space, which can result in unrealistic configurations or biased outcomes. We address this pitfall by introducing a Gaussian prior derived from the DiffVox vocal preset dataset over the parameter space. The resulting optimisation is equivalent to maximum-a-posteriori estimation. Evaluations on vocal effects transfer on the MedleyDB dataset show significant improvements across metrics compared to baselines, including a blind audio effects estimator, nearest-neighbour approaches, and uncalibrated ST-ITO. The proposed calibration reduces the parameter mean squared error by up to 33% and more closely matches the reference style. Subjective evaluations with 16 participants confirm the superiority of our method in limited data regimes. This work demonstrates how incorporating prior knowledge at inference time enhances audio effects transfer, paving the way for more effective and realistic audio processing systems.

comment: Published at WASPAA 2025

♻ ☆ Beat Tracking as Object Detection

Jaehoon Ahn, Moon-Ryul Jung

Recent beat and downbeat tracking models (e.g., RNNs, TCNs, Transformers) output frame-level activations. We propose reframing this task as object detection, where beats and downbeats are modeled as temporal "objects." Adapting the FCOS detector from computer vision to 1D audio, we replace its original backbone with WaveBeat's temporal feature extractor and add a Feature Pyramid Network to capture multi-scale temporal patterns. The model predicts overlapping beat/downbeat intervals with confidence scores, followed by non-maximum suppression (NMS) to select final predictions. This NMS step serves a similar role to DBNs in traditional trackers, but is simpler and less heuristic. Evaluated on standard music datasets, our approach achieves competitive results, showing that object detection techniques can effectively model musical beats with minimal adaptation.

comment: 11 pages, 4 figures, 5 tables

♻ ☆ MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction NeurIPS 2025

Yunkee Chae, Kyogu Lee

We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.

comment: Accepted by NeurIPS 2025

☆ Magnitude and Phase-based Feature Fusion Using Co-attention Mechanism for Speaker recognition

Rongfeng Su, Mengjie Du, Xiaokang Liu, Lan Wang, Nan Yan

Phase-based features related to vocal source characteristics can be incorporated into magnitude-based speaker recognition systems to improve the system performance. However, traditional feature-level fusion methods typically ignore the unique contributions of speaker semantics in the magnitude and phase domains. To address this issue, this paper proposed a feature-level fusion framework using the co-attention mechanism for speaker recognition. The framework consists of two separate sub-networks for the magnitude and phase domains respectively. Then, the intermediate high-level outputs of both domains are fused by the co-attention mechanism before a pooling layer. A correlation matrix from the co-attention module is supposed to re-assign the weights for dynamically scaling contributions in the magnitude and phase domains according to different pronunciations. Experiments on VoxCeleb showed that the proposed feature-level fusion strategy using the co-attention mechanism gave the Top-1 accuracy of 97.20%, outperforming the state-of-the-art system with 0.82% absolutely, and obtained EER reduction of 0.45% compared to single feature system using FBank.

☆ MC-LExt: Multi-Channel Target Speaker Extraction with Onset-Prompted Speaker Conditioning Mechanism

Tongtao Ling, Shulin He, Pengjie Shen, Zhong-Qiu Wang

Multi-channel target speaker extraction (MC-TSE) aims to extract a target speaker's voice from multi-speaker signals captured by multiple microphones. Existing methods often rely on auxiliary clues such as direction-of-arrival (DOA) or speaker embeddings. However, DOA-based approaches depend on explicit direction estimation and are sensitive to microphone array geometry, while methods based on speaker embeddings model speaker identity in an implicit manner and may degrade in noisy-reverberant conditions. To address these limitations, we propose multi-channel listen to extract (MC-LExt), a simple but highly-effective framework for MC-TSE. Our key idea is to prepend a short enrollment utterance of the target speaker to each channel of the multi-channel mixture, providing an onset-prompted conditioning signal that can guide TSE. This design allows the deep neural network (DNN) to learn spatial and speaker identity cues jointly in a fully end-to-end manner. Experiments on noisy-reverberant benchmarks, including WHAMR! and MC-Libri2Mix, demonstrate the effectiveness of MC-TSE.

comment: 5 pages, 2 figures

☆ Quantization-Based Score Calibration for Few-Shot Keyword Spotting with Dynamic Time Warping in Noisy Environments

Kevin Wilkinghoff, Alessia Cornaggia-Urrigshardt, Zheng-Hua Tan

Detecting occurrences of keywords with keyword spotting (KWS) systems requires thresholding continuous detection scores. Selecting appropriate thresholds is a non-trivial task, typically relying on optimizing the performance on a validation dataset. However, such greedy threshold selection often leads to suboptimal performance on unseen data, particularly in varying or noisy acoustic environments or few-shot settings. In this work, we investigate detection threshold estimation for template-based open-set few-shot KWS using dynamic time warping on noisy speech data. To mitigate the performance degradation caused by suboptimal thresholds, we propose a score calibration approach consisting of two different steps: quantizing embeddings and normalizing detection scores using the quantization error prior to thresholding. Experiments on KWS-DailyTalk with simulated high frequency radio channels show that the proposed calibration approach simplifies the choice of detection thresholds and significantly improves the resulting performance.

☆ Towards Blind Data Cleaning: A Case Study in Music Source Separation ICASSP 2026

Azalea Gui, Woosung Choi, Junghyun Koo, Kazuki Shimada, Takashi Shibuya, Joan Serrà, Wei-Hsiang Liao, Yuki Mitsufuji

The performance of deep learning models for music source separation heavily depends on training data quality. However, datasets are often corrupted by difficult-to-detect artifacts such as audio bleeding and label noise. Since the type and extent of contamination are typically unknown, cleaning methods targeting specific corruptions are often impractical. This paper proposes and evaluates two distinct, noise-agnostic data cleaning methods to address this challenge. The first approach uses data attribution via unlearning to identify and filter out training samples that contribute the least to producing clean outputs. The second leverages the Fr\'echet Audio Distance to measure and remove samples that are perceptually dissimilar to a small and trusted clean reference set. On a dataset contaminated with a simulated distribution of real-world noise, our unlearning-based methods produced a cleaned dataset and a corresponding model that outperforms both the original contaminated data and the small clean reference set used for cleaning. This result closes approximately 66.7\% of the performance gap between the contaminated baseline and a model trained on the same dataset without any contamination. Unlike methods tailored for specific artifacts, our noise-agnostic approaches offer a more generic and broadly applicable solution for curating high-quality training data.

comment: Submitted to IEEE ICASSP 2026

☆ DroneAudioset: An Audio Dataset for Drone-based Search and Rescue

Chitralekha Gupta, Soundarya Ramesh, Praveen Sasikumar, Kian Peen Yeo, Suranga Nanayakkara

Unmanned Aerial Vehicles (UAVs) or drones, are increasingly used in search and rescue missions to detect human presence. Existing systems primarily leverage vision-based methods which are prone to fail under low-visibility or occlusion. Drone-based audio perception offers promise but suffers from extreme ego-noise that masks sounds indicating human presence. Existing datasets are either limited in diversity or synthetic, lacking real acoustic interactions, and there are no standardized setups for drone audition. To this end, we present DroneAudioset (The dataset is publicly available at https://huggingface.co/datasets/ahlab-drone-project/DroneAudioSet/ under the MIT license), a comprehensive drone audition dataset featuring 23.5 hours of annotated recordings, covering a wide range of signal-to-noise ratios (SNRs) from -57.2 dB to -2.5 dB, across various drone types, throttles, microphone configurations as well as environments. The dataset enables development and systematic evaluation of noise suppression and classification methods for human-presence detection under challenging conditions, while also informing practical design considerations for drone audition systems, such as microphone placement trade-offs, and development of drone noise-aware audio processing. This dataset is an important step towards enabling design and deployment of drone-audition systems.

comment: Accepted in Neurips (Datasets and Benchmarks Track) 2025. The first two authors are equal contributors

☆ LDCodec: A high quality neural audio codec with low-complexity decoder

Jiawei Jiang, Linping Xu, Dejun Zhang, Qingbo Huang, Xianjun Xia, Yijian Xiao

Neural audio coding has been shown to outperform classical audio coding at extremely low bitrates. However, the practical application of neural audio codecs is still limited by their elevated complexity. To address this challenge, we have developed a high-quality neural audio codec with a low-complexity decoder, named LDCodec (Low-complexity Decoder Neural Audio Codec), specifically designed for on-demand streaming media clients, such as smartphones. Specifically, we introduced a novel residual unit combined with Long-term and Short-term Residual Vector Quantization (LSRVQ), subband-fullband frequency discriminators, and perceptual loss functions. This combination results in high-quality audio reconstruction with lower complexity. Both our subjective and objective tests demonstrated that our proposed LDCodec at 6kbps outperforms Opus at 12kbps.

☆ Extending Audio Context for Long-Form Understanding in Large Audio-Language Models

Yuatyong Chaichana, Pittawat Taveekitworachai, Warit Sirichotedumrong, Potsawee Manakul, Kunat Pipatanakul

Large Audio-Language Models (LALMs) are often constrained by short audio context windows, even when their text backbones support long contexts, limiting long-form audio understanding. Prior work has introduced context-extension methods (e.g. YaRN) on unimodal LLMs, yet their application to LALMs remains unexplored. First, building on RoPE-based context extension, we introduce Partial YaRN, a training-free, audio-only extension method that modifies only audio token positions, leaving text positions intact to preserve the base LLM's text capabilities. Second, we propose Virtual Longform Audio Training (VLAT), a training strategy that extends Partial YaRN into a training-time positional augmentation. VLAT simulates diverse audio lengths during training, enabling generalization to inputs far longer than those seen in training and improving robustness for long-context audio understanding. Our experiments on SALMONN and Qwen2-Audio show that Partial YaRN outperforms the original models across wide range of settings, and VLAT training strategy provides substantial improvement, achieving strong performance on long audio of unseen lengths.

☆ LongCat-Audio-Codec: An Audio Tokenizer and Detokenizer Solution Designed for Speech Large Language Models

Xiaohan Zhao, Hongyu Xiang, Shengze Ye, Song Li, Zhengkun Tian, Guanyu Chen, Ke Ding, Guanglu Wan

This paper presents LongCat-Audio-Codec, an audio tokenizer and detokenizer solution designed for industrial grade end-to-end speech large language models. By leveraging a decoupled model architecture and a multistage training strategy, LongCat-Audio-Codec exhibits robust semantic modeling capabilities, flexible acoustic feature extraction capabilities, and low-latency streaming synthesis capabilities. It encodes speech at an ultra-low frame rate of 16.67 Hz, with a minimum bitrate of 0.43 kbps and a maximum bitrate of 0.87 kbps. Evaluation results demonstrate that LongCat-Audio-Codec achieves strong speech intelligibility and is capable of synthesizing highquality speech at low bitrate, thus effectively balancing coding efficiency and decoding quality. The inference code and model checkpoints of LongCat-Audio-Codec are available at: https://github.com/meituan-longcat/LongCat-Audio-Codec.

☆ AsyncVoice Agent: Real-Time Explanation for LLM Planning and Reasoning ASRU 2025

Yueqian Lin, Zhengmian Hu, Jayakumar Subramanian, Qinsi Wang, Nikos Vlassis, Hai "Helen" Li, Yiran Chen

Effective human-AI collaboration on complex reasoning tasks requires that users understand and interact with the model's process, not just receive an output. However, the monolithic text from methods like Chain-of-Thought (CoT) prevents this, as current interfaces lack real-time verbalization and robust user barge-in. We present AsyncVoice Agent, a system whose asynchronous architecture decouples a streaming LLM backend from a conversational voice frontend. This design allows narration and inference to run in parallel, empowering users to interrupt, query, and steer the model's reasoning process at any time. Objective benchmarks show this approach reduces interaction latency by more than 600x compared to monolithic baselines while ensuring high fidelity and competitive task accuracy. By enabling a two-way dialogue with a model's thought process, AsyncVoice Agent offers a new paradigm for building more effective, steerable, and trustworthy human-AI systems for high-stakes tasks.

comment: Accepted to the IEEE ASRU 2025 Demo Track

♻ ☆ Summarizing Speech: A Comprehensive Survey EMNLP 2025

Fabian Retkowski, Maike Züfle, Andreas Sudmann, Dinah Pfau, Shinji Watanabe, Jan Niehues, Alexander Waibel

Speech summarization has become an essential tool for efficiently managing and accessing the growing volume of spoken and audiovisual content. However, despite its increasing importance, speech summarization remains loosely defined. The field intersects with several research areas, including speech recognition, text summarization, and specific applications like meeting summarization. This survey not only examines existing datasets and evaluation protocols, which are crucial for assessing the quality of summarization approaches, but also synthesizes recent developments in the field, highlighting the shift from traditional systems to advanced models like fine-tuned cascaded architectures and end-to-end solutions. In doing so, we surface the ongoing challenges, such as the need for realistic evaluation benchmarks, multilingual datasets, and long-context handling.

comment: Accepted to EMNLP 2025

♻ ☆ Benchmarking Fake Voice Detection in the Fake Voice Generation Arms Race

Xutao Mao, Ke Li, Cameron Baird, Ezra Xuanru Tao, Dan Lin

The rapid advancement of fake voice generation technology has ignited a race with detection systems, creating an urgent need to secure the audio ecosystem. However, existing benchmarks suffer from a critical limitation: they typically aggregate diverse fake voice samples into a single dataset for evaluation. This practice masks method-specific artifacts and obscures the varying performance of detectors against different generation paradigms, preventing a nuanced understanding of their true vulnerabilities. To address this gap, we introduce the first ecosystem-level benchmark that systematically evaluates the interplay between 17 state-of-the-art fake voice generators and 8 leading detectors through a novel one-to-one evaluation protocol. This fine-grained analysis exposes previously hidden vulnerabilities and sensitivities that are missed by traditional aggregated testing. We also propose unified scoring systems to quantify both the evasiveness of generators and the robustness of detectors, enabling fair and direct comparisons. Our extensive cross-domain evaluation reveals that modern generators, particularly those based on neural audio codecs and flow matching, consistently evade top-tier detectors. We found that no single detector is universally robust; their effectiveness varies dramatically depending on the generator's architecture, highlighting a significant generalization gap in current defenses. This work provides a more realistic assessment of the threat landscape and offers actionable insights for building the next generation of detection systems.

♻ ☆ BandCondiNet: Parallel Transformers-based Conditional Popular Music Generation with Multi-View Features

Jing Luo, Xinyu Yang, Dorien Herremans

Conditional music generation offers significant advantages in terms of user convenience and control, presenting great potential in AI-generated content research. However, building conditional generative systems for multitrack popular songs presents three primary challenges: insufficient fidelity of input conditions, poor structural modeling, and inadequate inter-track harmony learning in generative models. To address these issues, we propose BandCondiNet, a conditional model based on parallel Transformers, designed to process the multiple music sequences and generate high-quality multitrack samples. Specifically, we propose multi-view features across time and instruments as high-fidelity conditions. Moreover, we propose two specialized modules for BandCondiNet: Structure Enhanced Attention (SEA) to strengthen the musical structure, and Cross-Track Transformer (CTT) to enhance inter-track harmony. We conducted both objective and subjective evaluations on two popular music datasets with different sequence lengths. Objective results on the shorter dataset show that BandCondiNet outperforms other conditional models in 9 out of 10 metrics related to fidelity and inference speed, with the exception of Chord Accuracy. On the longer dataset, BandCondiNet surpasses all conditional models across all 10 metrics. Subjective evaluations across four criteria reveal that BandCondiNet trained on the shorter dataset performs best in Richness and performs comparably to state-of-the-art models in the other three criteria, while significantly outperforming them across all criteria when trained on the longer dataset. To further expand the application scope of BandCondiNet, future work should focus on developing an advanced conditional model capable of adapting to more user-friendly input conditions and supporting flexible instrumentation.

comment: To appear in ESWA. Demo page: https://chinglohsiu.github.io/files/bandcondinet.html

♻ ☆ EmoSphere-SER: Enhancing Speech Emotion Recognition Through Spherical Representation with Auxiliary Classification

Deok-Hyeon Cho, Hyung-Seok Oh, Seung-Bin Kim, Seong-Whan Lee

Speech emotion recognition predicts a speaker's emotional state from speech signals using discrete labels or continuous dimensions such as arousal, valence, and dominance (VAD). We propose EmoSphere-SER, a joint model that integrates spherical VAD region classification to guide VAD regression for improved emotion prediction. In our framework, VAD values are transformed into spherical coordinates that are divided into multiple spherical regions, and an auxiliary classification task predicts which spherical region each point belongs to, guiding the regression process. Additionally, we incorporate a dynamic weighting scheme and a style pooling layer with multi-head self-attention to capture spectral and temporal dynamics, further boosting performance. This combined training strategy reinforces structured learning and improves prediction consistency. Experimental results show that our approach exceeds baseline methods, confirming the validity of the proposed framework.

comment: Proceedings of Interspeech 2025

♻ ☆ DiEmo-TTS: Disentangled Emotion Representations via Self-Supervised Distillation for Cross-Speaker Emotion Transfer in Text-to-Speech

Deok-Hyeon Cho, Hyung-Seok Oh, Seung-Bin Kim, Seong-Whan Lee

Cross-speaker emotion transfer in speech synthesis relies on extracting speaker-independent emotion embeddings for accurate emotion modeling without retaining speaker traits. However, existing timbre compression methods fail to fully separate speaker and emotion characteristics, causing speaker leakage and degraded synthesis quality. To address this, we propose DiEmo-TTS, a self-supervised distillation method to minimize emotional information loss and preserve speaker identity. We introduce cluster-driven sampling and information perturbation to preserve emotion while removing irrelevant factors. To facilitate this process, we propose an emotion clustering and matching approach using emotional attribute prediction and speaker embeddings, enabling generalization to unlabeled data. Additionally, we designed a dual conditioning transformer to integrate style features better. Experimental results confirm the effectiveness of our method in learning speaker-irrelevant emotion embeddings.

comment: Proceedings of Interspeech 2025

♻ ☆ Improving Inference-Time Optimisation for Vocal Effects Style Transfer with a Gaussian Prior SP

Chin-Yun Yu, Marco A. Martínez-Ramírez, Junghyun Koo, Wei-Hsiang Liao, Yuki Mitsufuji, György Fazekas

Style Transfer with Inference-Time Optimisation (ST-ITO) is a recent approach for transferring the applied effects of a reference audio to an audio track. It optimises the effect parameters to minimise the distance between the style embeddings of the processed audio and the reference. However, this method treats all possible configurations equally and relies solely on the embedding space, which can result in unrealistic configurations or biased outcomes. We address this pitfall by introducing a Gaussian prior derived from the DiffVox vocal preset dataset over the parameter space. The resulting optimisation is equivalent to maximum-a-posteriori estimation. Evaluations on vocal effects transfer on the MedleyDB dataset show significant improvements across metrics compared to baselines, including a blind audio effects estimator, nearest-neighbour approaches, and uncalibrated ST-ITO. The proposed calibration reduces the parameter mean squared error by up to 33% and more closely matches the reference style. Subjective evaluations with 16 participants confirm the superiority of our method in limited data regimes. This work demonstrates how incorporating prior knowledge at inference time enhances audio effects transfer, paving the way for more effective and realistic audio processing systems.

comment: Published at WASPAA 2025

♻ ☆ MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction NeurIPS 2025

Yunkee Chae, Kyogu Lee

We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.

comment: Accepted by NeurIPS 2025

☆ TASLA: Text-Aligned Speech Tokens with Multiple Layer-Aggregation

Ming-Hao Hsu, Liang-Hsuan Tseng, Hung-yi Lee, Zhizheng Wu

We propose Text-Aligned Speech Tokens with Multiple Layer-Aggregation (TASLA), which is a text-aligned speech tokenization framework that aims to address the problem that under a low-frame-rate and text-aligned regime, single-source speech tokens may lose acoustic details during reconstruction. On the other hand, this paper further explains how different encoder layers collaborate to capture comprehensive acoustic features for tokenization. Previous work, TASTE, proposed the text-aligned speech tokenization framework, which is a LM-friendly architecture, but struggles to capture acoustic details. We address this trade-off with two components: Multi-Layer Dynamic Attention (MLDA), which lets each text position adaptively mix shallow/deep features from a frozen speech encoder, and Finite Scalar Quantization (FSQ), a simple per-dimension discretization with smooth optimization. At about 2.62 Hz (tokens/s), TASLA consistently improves prosody and achieves competitive quality over TASTE on in-domain (LibriSpeech) and OOD (EXPRESSO, Voxceleb) sets. We further demonstrate that dynamic layer mixing is correlated with spectral flux and explains why MLDA preserves prosody under a low frame rate with extreme feature compression.

☆ Sound Masking Strategies for Interference with Mosquito Hearing

Justin Faber, Alexandros C Alampounti, Marcos Georgiades, Joerg T Albert, Dolores Bozovic

The use of auditory masking has long been of interest in psychoacoustics and for engineering purposes, in order to cover sounds that are disruptive to humans or to species whose habitats overlap with ours. In most cases, we seek to minimize the disturbances to the communication of wildlife. However, in the case of pathogen-carrying insects, we may want to maximize these disturbances as a way to control populations. In the current work, we explore candidate masking strategies for a generic model of active auditory systems and a model of the mosquito auditory system. For both models, we find that masks with all acoustic power focused into just one or a few frequencies perform best. We propose that masks based on rapid frequency modulation are most effective for maximal disruption of information transfer and minimizing intelligibility. We hope that these results will serve to guide the avoidance or selection of possible acoustic signals for, respectively, maximizing or minimizing communication.

☆ If You Hold Me Without Hurting Me: Pathways to Designing Game Audio for Healthy Escapism and Player Well-being

Caio Nunes, Bosco Borges, Georgia Cruz, Ticianne Darin

Escapism in games can support recovery or lead to harmful avoidance. Self-regulation, understood as combining autonomy with positive outcomes, is key to this distinction. We argue that audio, often overlooked, plays a central role in regulation. It can modulate arousal, mark transitions, and provide closure, yet its contribution to well-being remains underexplored. This paper identifies methodological and accessibility gaps that limit recognition of audio's potential and outlines ways to address them. We aim to encourage researchers and developers to integrate audio more deliberately into the design and study of healthier escapist play.

comment: 5 pages. Presented and discussed at the CHI PLAY 2025 Workshop Exploring Future Directions for Healthy Escapism and Self-Regulation in Games, Pittsburgh, USA, October 13, 2025

☆ SpeechLLM-as-Judges: Towards General and Interpretable Speech Quality Evaluation

Hui Wang, Jinghua Zhao, Yifan Yang, Shujie Liu, Junyang Chen, Yanzhe Zhang, Shiwan Zhao, Jinyu Li, Jiaming Zhou, Haoqin Sun, Yan Lu, Yong Qin

Generative speech technologies are progressing rapidly, but evaluating the perceptual quality of synthetic speech remains a core challenge. Existing methods typically rely on scalar scores or binary decisions, which lack interpretability and generalization across tasks and languages. We present SpeechLLM-as-Judges, a new paradigm for enabling large language models (LLMs) to conduct structured and explanation-based speech quality evaluation. To support this direction, we introduce SpeechEval, a large-scale dataset containing 32,207 multilingual speech clips and 128,754 annotations spanning four tasks: quality assessment, pairwise comparison, improvement suggestion, and deepfake detection. Based on this resource, we develop SQ-LLM, a speech-quality-aware LLM trained with chain-of-thought reasoning and reward optimization to improve capability. Experimental results show that SQ-LLM delivers strong performance across tasks and languages, revealing the potential of this paradigm for advancing speech quality evaluation. Relevant resources will be open-sourced.

☆ AudioEval: Automatic Dual-Perspective and Multi-Dimensional Evaluation of Text-to-Audio-Generation

Hui Wang, Jinghua Zhao, Cheng Liu, Yuhang Jia, Haoqin Sun, Jiaming Zhou, Yong Qin

Text-to-audio (TTA) is rapidly advancing, with broad potential in virtual reality, accessibility, and creative media. However, evaluating TTA quality remains difficult: human ratings are costly and limited, while existing objective metrics capture only partial aspects of perceptual quality. To address this gap, we introduce AudioEval, the first large-scale TTA evaluation dataset, containing 4,200 audio samples from 24 systems with 126,000 ratings across five perceptual dimensions, annotated by both experts and non-experts. Based on this resource, we propose Qwen-DisQA, a multimodal scoring model that jointly processes text prompts and generated audio to predict human-like quality ratings. Experiments show its effectiveness in providing reliable and scalable evaluation. The dataset will be made publicly available to accelerate future research.

☆ Big Data Approaches to Bovine Bioacoustics: A FAIR-Compliant Dataset and Scalable ML Framework for Precision Livestock Welfare

Mayuri Kate, Suresh Neethirajan

The convergence of IoT sensing, edge computing, and machine learning is transforming precision livestock farming. Yet bioacoustic data streams remain underused because of computational complexity and ecological validity challenges. We present one of the most comprehensive bovine vocalization datasets to date, with 569 curated clips covering 48 behavioral classes, recorded across three commercial dairy farms using multiple microphone arrays and expanded to 2900 samples through domain informed augmentation. This FAIR compliant resource addresses major Big Data challenges - volume (90 hours of recordings, 65.6 GB), variety (multi farm and multi zone acoustics), velocity (real time processing), and veracity (noise robust feature extraction). Our distributed processing framework integrates advanced denoising using iZotope RX, multimodal synchronization through audio and video alignment, and standardized feature engineering with 24 acoustic descriptors generated from Praat, librosa, and openSMILE. Preliminary benchmarks reveal distinct class level acoustic patterns for estrus detection, distress classification, and maternal communication. The datasets ecological realism, reflecting authentic barn acoustics rather than controlled settings, ensures readiness for field deployment. This work establishes a foundation for animal centered AI, where bioacoustic data enable continuous and non invasive welfare assessment at industrial scale. By releasing standardized pipelines and detailed metadata, we promote reproducible research that connects Big Data analytics, sustainable agriculture, and precision livestock management. The framework supports UN SDG 9, showing how data science can turn traditional farming into intelligent, welfare optimized systems that meet global food needs while upholding ethical animal care.

comment: 40 pages, 14 figures, 9 Tables

☆ Revisit Modality Imbalance at the Decision Layer

Xiaoyu Ma, Hao Chen

Multimodal learning integrates information from different modalities to enhance model performance, yet it often suffers from modality imbalance, where dominant modalities overshadow weaker ones during joint optimization. This paper reveals that such an imbalance not only occurs during representation learning but also manifests significantly at the decision layer. Experiments on audio-visual datasets (CREMAD and Kinetic-Sounds) show that even after extensive pretraining and balanced optimization, models still exhibit systematic bias toward certain modalities, such as audio. Further analysis demonstrates that this bias originates from intrinsic disparities in feature-space and decision-weight distributions rather than from optimization dynamics alone. We argue that aggregating uncalibrated modality outputs at the fusion stage leads to biased decision-layer weighting, hindering weaker modalities from contributing effectively. To address this, we propose that future multimodal systems should focus more on incorporate adaptive weight allocation mechanisms at the decision layer, enabling relative balanced according to the capabilities of each modality.

comment: Some Insights in Balanced Multimodal Learning

☆ Beat Detection as Object Detection

Jaehoon Ahn, Moon-Ryul Jung

Recent beat and downbeat tracking models (e.g., RNNs, TCNs, Transformers) output frame-level activations. We propose reframing this task as object detection, where beats and downbeats are modeled as temporal "objects." Adapting the FCOS detector from computer vision to 1D audio, we replace its original backbone with WaveBeat's temporal feature extractor and add a Feature Pyramid Network to capture multi-scale temporal patterns. The model predicts overlapping beat/downbeat intervals with confidence scores, followed by non-maximum suppression (NMS) to select final predictions. This NMS step serves a similar role to DBNs in traditional trackers, but is simpler and less heuristic. Evaluated on standard music datasets, our approach achieves competitive results, showing that object detection techniques can effectively model musical beats with minimal adaptation.

comment: 11 pages, 4 figures, 5 tables

☆ Do Joint Language-Audio Embeddings Encode Perceptual Timbre Semantics?

Qixin Deng, Bryan Pardo, Thrasyvoulos N Pappas

Understanding and modeling the relationship between language and sound is critical for applications such as music information retrieval,text-guided music generation, and audio captioning. Central to these tasks is the use of joint language-audio embedding spaces, which map textual descriptions and auditory content into a shared embedding space. While multimodal embedding models such as MS-CLAP, LAION-CLAP, and MuQ-MuLan have shown strong performance in aligning language and audio, their correspondence to human perception of timbre, a multifaceted attribute encompassing qualities such as brightness, roughness, and warmth, remains underexplored. In this paper, we evaluate the above three joint language-audio embedding models on their ability to capture perceptual dimensions of timbre. Our findings show that LAION-CLAP consistently provides the most reliable alignment with human-perceived timbre semantics across both instrumental sounds and audio effects.

☆ TRI-DEP: A Trimodal Comparative Study for Depression Detection Using Speech, Text, and EEG

Annisaa Fitri Nurfidausi, Eleonora Mancini, Paolo Torroni

Depression is a widespread mental health disorder, yet its automatic detection remains challenging. Prior work has explored unimodal and multimodal approaches, with multimodal systems showing promise by leveraging complementary signals. However, existing studies are limited in scope, lack systematic comparisons of features, and suffer from inconsistent evaluation protocols. We address these gaps by systematically exploring feature representations and modelling strategies across EEG, together with speech and text. We evaluate handcrafted features versus pre-trained embeddings, assess the effectiveness of different neural encoders, compare unimodal, bimodal, and trimodal configurations, and analyse fusion strategies with attention to the role of EEG. Consistent subject-independent splits are applied to ensure robust, reproducible benchmarking. Our results show that (i) the combination of EEG, speech and text modalities enhances multimodal detection, (ii) pretrained embeddings outperform handcrafted features, and (iii) carefully designed trimodal models achieve state-of-the-art performance. Our work lays the groundwork for future research in multimodal depression detection.

☆ SpeechLLM-as-Judges: Towards General and Interpretable Speech Quality Evaluation

Hui Wang, Jinghua Zhao, Yifan Yang, Shujie Liu, Junyang Chen, Yanzhe Zhang, Shiwan Zhao, Jinyu Li, Jiaming Zhou, Haoqin Sun, Yan Lu, Yong Qin

Generative speech technologies are progressing rapidly, but evaluating the perceptual quality of synthetic speech remains a core challenge. Existing methods typically rely on scalar scores or binary decisions, which lack interpretability and generalization across tasks and languages. We present SpeechLLM-as-Judges, a new paradigm for enabling large language models (LLMs) to conduct structured and explanation-based speech quality evaluation. To support this direction, we introduce SpeechEval, a large-scale dataset containing 32,207 multilingual speech clips and 128,754 annotations spanning four tasks: quality assessment, pairwise comparison, improvement suggestion, and deepfake detection. Based on this resource, we develop SQ-LLM, a speech-quality-aware LLM trained with chain-of-thought reasoning and reward optimization to improve capability. Experimental results show that SQ-LLM delivers strong performance across tasks and languages, revealing the potential of this paradigm for advancing speech quality evaluation. Relevant resources will be open-sourced.

☆ AudioEval: Automatic Dual-Perspective and Multi-Dimensional Evaluation of Text-to-Audio-Generation

Hui Wang, Jinghua Zhao, Cheng Liu, Yuhang Jia, Haoqin Sun, Jiaming Zhou, Yong Qin

Text-to-audio (TTA) is rapidly advancing, with broad potential in virtual reality, accessibility, and creative media. However, evaluating TTA quality remains difficult: human ratings are costly and limited, while existing objective metrics capture only partial aspects of perceptual quality. To address this gap, we introduce AudioEval, the first large-scale TTA evaluation dataset, containing 4,200 audio samples from 24 systems with 126,000 ratings across five perceptual dimensions, annotated by both experts and non-experts. Based on this resource, we propose Qwen-DisQA, a multimodal scoring model that jointly processes text prompts and generated audio to predict human-like quality ratings. Experiments show its effectiveness in providing reliable and scalable evaluation. The dataset will be made publicly available to accelerate future research.

☆ Spatially Aware Self-Supervised Models for Multi-Channel Neural Speaker Diarization ICASSP 2026

Jiangyu Han, Ruoyu Wang, Yoshiki Masuyama, Marc Delcroix, Johan Rohdin, Jun Du, Lukas Burget

Self-supervised models such as WavLM have demonstrated strong performance for neural speaker diarization. However, these models are typically pre-trained on single-channel recordings, limiting their effectiveness in multi-channel scenarios. Existing diarization systems built on these models often rely on DOVER-Lap to combine outputs from individual channels. Although effective, this approach incurs substantial computational overhead and fails to fully exploit spatial information. In this work, building on DiariZen, a pipeline that combines WavLM-based local endto-end neural diarization with speaker embedding clustering, we introduce a lightweight approach to make pre-trained WavLM spatially aware by inserting channel communication modules into the early layers. Our method is agnostic to both the number of microphone channels and array topologies, ensuring broad applicability. We further propose to fuse multi-channel speaker embeddings by leveraging spatial attention weights. Evaluations on five public datasets show consistent improvements over single-channel baselines and demonstrate superior performance and efficiency compared with DOVER-Lap. Our source code is publicly available at https://github.com/BUTSpeechFIT/DiariZen.

comment: Submitted to ICASSP 2026

☆ Big Data Approaches to Bovine Bioacoustics: A FAIR-Compliant Dataset and Scalable ML Framework for Precision Livestock Welfare

Mayuri Kate, Suresh Neethirajan

The convergence of IoT sensing, edge computing, and machine learning is transforming precision livestock farming. Yet bioacoustic data streams remain underused because of computational complexity and ecological validity challenges. We present one of the most comprehensive bovine vocalization datasets to date, with 569 curated clips covering 48 behavioral classes, recorded across three commercial dairy farms using multiple microphone arrays and expanded to 2900 samples through domain informed augmentation. This FAIR compliant resource addresses major Big Data challenges - volume (90 hours of recordings, 65.6 GB), variety (multi farm and multi zone acoustics), velocity (real time processing), and veracity (noise robust feature extraction). Our distributed processing framework integrates advanced denoising using iZotope RX, multimodal synchronization through audio and video alignment, and standardized feature engineering with 24 acoustic descriptors generated from Praat, librosa, and openSMILE. Preliminary benchmarks reveal distinct class level acoustic patterns for estrus detection, distress classification, and maternal communication. The datasets ecological realism, reflecting authentic barn acoustics rather than controlled settings, ensures readiness for field deployment. This work establishes a foundation for animal centered AI, where bioacoustic data enable continuous and non invasive welfare assessment at industrial scale. By releasing standardized pipelines and detailed metadata, we promote reproducible research that connects Big Data analytics, sustainable agriculture, and precision livestock management. The framework supports UN SDG 9, showing how data science can turn traditional farming into intelligent, welfare optimized systems that meet global food needs while upholding ethical animal care.

comment: 40 pages, 14 figures, 9 Tables

☆ Revisit Modality Imbalance at the Decision Layer

Xiaoyu Ma, Hao Chen

Multimodal learning integrates information from different modalities to enhance model performance, yet it often suffers from modality imbalance, where dominant modalities overshadow weaker ones during joint optimization. This paper reveals that such an imbalance not only occurs during representation learning but also manifests significantly at the decision layer. Experiments on audio-visual datasets (CREMAD and Kinetic-Sounds) show that even after extensive pretraining and balanced optimization, models still exhibit systematic bias toward certain modalities, such as audio. Further analysis demonstrates that this bias originates from intrinsic disparities in feature-space and decision-weight distributions rather than from optimization dynamics alone. We argue that aggregating uncalibrated modality outputs at the fusion stage leads to biased decision-layer weighting, hindering weaker modalities from contributing effectively. To address this, we propose that future multimodal systems should focus more on incorporate adaptive weight allocation mechanisms at the decision layer, enabling relative balanced according to the capabilities of each modality.

comment: Some Insights in Balanced Multimodal Learning

☆ A Robust Classification Method using Hybrid Word Embedding for Early Diagnosis of Alzheimer's Disease

Yangyang Li

Early detection of Alzheimer's Disease (AD) is greatly beneficial to AD patients, leading to early treatments that lessen symptoms and alleviating financial burden of health care. As one of the leading signs of AD, language capability changes can be used for early diagnosis of AD. In this paper, I develop a robust classification method using hybrid word embedding and fine-tuned hyperparameters to achieve state-of-the-art accuracy in the early detection of AD. Specifically, we create a hybrid word embedding based on word vectors from Doc2Vec and ELMo to obtain perplexity scores of the sentences. The scores identify whether a sentence is fluent or not and capture semantic context of the sentences. I enrich the word embedding by adding linguistic features to analyze syntax and semantics. Further, we input an embedded feature vector into logistic regression and fine tune hyperparameters throughout the pipeline. By tuning hyperparameters of the machine learning pipeline (e.g., model regularization parameter, learning rate and vector size of Doc2Vec, and vector size of ELMo), I achieve 91% classification accuracy and an Area Under the Curve (AUC) of 97% in distinguishing early AD from healthy subjects. Based on my knowledge, my model with 91% accuracy and 97% AUC outperforms the best existing NLP model for AD diagnosis with an accuracy of 88% [32]. I study the model stability through repeated experiments and find that the model is stable even though the training data is split randomly (standard deviation of accuracy = 0.0403; standard deviation of AUC = 0.0174). This affirms our proposed method is accurate and stable. This model can be used as a large-scale screening method for AD, as well as a complementary examination for doctors to detect AD.

comment: Peer-reviewed and published in Proceedings of the 2020 3rd International Conference on Algorithms, Computing and Artificial Intelligence (ACAI 2020). 7 pages, 5 figures

☆ Do Joint Language-Audio Embeddings Encode Perceptual Timbre Semantics?

Qixin Deng, Bryan Pardo, Thrasyvoulos N Pappas

Understanding and modeling the relationship between language and sound is critical for applications such as music information retrieval,text-guided music generation, and audio captioning. Central to these tasks is the use of joint language-audio embedding spaces, which map textual descriptions and auditory content into a shared embedding space. While multimodal embedding models such as MS-CLAP, LAION-CLAP, and MuQ-MuLan have shown strong performance in aligning language and audio, their correspondence to human perception of timbre, a multifaceted attribute encompassing qualities such as brightness, roughness, and warmth, remains underexplored. In this paper, we evaluate the above three joint language-audio embedding models on their ability to capture perceptual dimensions of timbre. Our findings show that LAION-CLAP consistently provides the most reliable alignment with human-perceived timbre semantics across both instrumental sounds and audio effects.

♻ ☆ Pinhole Effect on Linkability and Dispersion in Speaker Anonymization

Kong Aik Lee, Zeyan Liu, Liping Chen, Zhenhua Ling

Speaker anonymization aims to conceal speaker-specific attributes in speech signals, making the anonymized speech unlinkable to the original speaker identity. Recent approaches achieve this by disentangling speech into content and speaker components, replacing the latter with pseudo speakers. The anonymized speech can be mapped either to a common pseudo speaker shared across utterances or to distinct pseudo speakers unique to each utterance. This paper investigates the impact of these mapping strategies on three key dimensions: speaker linkability, dispersion in the anonymized speaker space, and de-identification from the original identity. Our findings show that using distinct pseudo speakers increases speaker dispersion and reduces linkability compared to common pseudo-speaker mapping, thereby enhancing privacy preservation. These observations are interpreted through the proposed pinhole effect, a conceptual framework introduced to explain the relationship between mapping strategies and anonymization performance. The hypothesis is validated through empirical evaluation.

comment: 6 pages, 2 figures

♻ ☆ SPIRIT: Patching Speech Language Models against Jailbreak Attacks

Amirbek Djanibekov, Nurdaulet Mukhituly, Kentaro Inui, Hanan Aldarmaki, Nils Lukas

Speech Language Models (SLMs) enable natural interactions via spoken instructions, which more effectively capture user intent by detecting nuances in speech. The richer speech signal introduces new security risks compared to text-based models, as adversaries can better bypass safety mechanisms by injecting imperceptible noise to speech. We analyze adversarial attacks and find that SLMs are substantially more vulnerable to jailbreak attacks, which can achieve a perfect 100% attack success rate in some instances. To improve security, we propose post-hoc patching defenses used to intervene during inference by modifying the SLM's activations that improve robustness up to 99% with (i) negligible impact on utility and (ii) without any re-training. We conduct ablation studies to maximize the efficacy of our defenses and improve the utility/security trade-off, validated with large-scale benchmarks unique to SLMs.

♻ ☆ Towards Inclusive Communication: A Unified Framework for Generating Spoken Language from Sign, Lip, and Audio

Jeong Hun Yeo, Hyeongseop Rha, Sungjune Park, Junil Won, Yong Man Ro

Audio is the primary modality for human communication and has driven the success of Automatic Speech Recognition (ASR) technologies. However, such audio-centric systems inherently exclude individuals who are deaf or hard of hearing. Visual alternatives such as sign language and lip reading offer effective substitutes, and recent advances in Sign Language Translation (SLT) and Visual Speech Recognition (VSR) have improved audio-less communication. Yet, these modalities have largely been studied in isolation, and their integration within a unified framework remains underexplored. In this paper, we propose the first unified framework capable of handling diverse combinations of sign language, lip movements, and audio for spoken-language text generation. We focus on three main objectives: (i) designing a unified, modality-agnostic architecture capable of effectively processing heterogeneous inputs; (ii) exploring the underexamined synergy among modalities, particularly the role of lip movements as non-manual cues in sign language comprehension; and (iii) achieving performance on par with or superior to state-of-the-art models specialized for individual tasks. Building on this framework, we achieve performance on par with or better than task-specific state-of-the-art models across SLT, VSR, ASR, and Audio-Visual Speech Recognition. Furthermore, our analysis reveals a key linguistic insight: explicitly modeling lip movements as a distinct modality significantly improves SLT performance by capturing critical non-manual cues.

♻ ☆ Non-invasive electromyographic speech neuroprosthesis: a geometric perspective

Harshavardhana T. Gowda, Lee M. Miller

We present a high-bandwidth, egocentric neuromuscular speech interface that translates $silently$ voiced articulations directly into text. We record surface electromyographic (EMG) signals from multiple articulatory sites on the face and neck as participants $silently$ articulate speech, enabling direct EMG-to-text translation. Such an interface has the potential to restore communication for individuals who have lost the ability to produce intelligible speech due to laryngectomy, neuromuscular disease, stroke, or trauma-induced damage (e.g., radiotherapy toxicity) to the speech articulators. Prior work has largely focused on mapping EMG collected during $audible$ articulation to time-aligned audio targets or transferring these targets to $silent$ EMG recordings, which inherently requires audio and limits applicability to patients who can no longer speak. In contrast, we propose an efficient representation of high-dimensional EMG signals and demonstrate direct sequence-to-sequence EMG-to-text conversion at the phonemic level without relying on time-aligned audio. All data, code, and model checkpoints are open-sourced at The dataset and code are available at: https://github.com/HarshavardhanaTG/emg2speech .

♻ ☆ A New Time Series Similarity Measure and Its Smart Grid Applications

Rui Yuan, Hossein Ranjbar, S. Ali Pourmousavi, Wen L. Soong, Andrew J. Black, Jon A. R. Liisberg, Julian Lemos-Vinasco

Many smart grid applications involve data mining, clustering, classification, identification, and anomaly detection, among others. These applications primarily depend on the measurement of similarity, which is the distance between different time series or subsequences of a time series. The commonly used time series distance measures, namely Euclidean Distance (ED) and Dynamic Time Warping (DTW), do not quantify the flexible nature of electricity usage data in terms of temporal dynamics. As a result, there is a need for a new distance measure that can quantify both the amplitude and temporal changes of electricity time series for smart grid applications, e.g., demand response and load profiling. This paper introduces a novel distance measure to compare electricity usage patterns. The method consists of two phases that quantify the effort required to reshape one time series into another, considering both amplitude and temporal changes. The proposed method is evaluated against ED and DTW using real-world data in three smart grid applications. Overall, the proposed measure outperforms ED and DTW in accurately identifying the best load scheduling strategy, anomalous days with irregular electricity usage, and determining electricity users' behind-the-meter (BTM) equipment.

comment: 6 pages, 5 figures conference

☆ Musical consonance: a review of theory and evidence on perception and preference of auditory roughness in humans and other animals

John M. McBride

The origins of consonance in human music has long been contested, and today there are three primary hypotheses: aversion to roughness, preference for harmonicity, and learned preferences from cultural exposure. While the evidence is currently insufficient to disentangle the contributions of these hypotheses, I propose several reasons why roughness is an especially promising area for future study. The aim of this review is to summarize and critically evaluate roughness theory and models, experimental data, to highlight areas that deserve further research. I identify 2 key areas: There are fundamental issues with the definition and interpretation of results due to tautology in the definition of roughness, and the lack of independence in empirical measurements. Despite extensive model development, there are many duplications and models have issues with data quality and overfitting. Future theory development should aim for model simplicity, and extra assumptions, features and parameters should be evaluated systematically. Model evaluation should aim to maximise the breadth of stimuli that are predicted.

☆ Closing the Gap Between Text and Speech Understanding in LLMs

Santiago Cuervo, Skyler Seto, Maureen de Seyssel, Richard He Bai, Zijin Gu, Tatiana Likhomanenko, Navdeep Jaitly, Zakaria Aldeneh

Large Language Models (LLMs) can be adapted to extend their text capabilities to speech inputs. However, these speech-adapted LLMs consistently underperform their text-based counterparts--and even cascaded pipelines--on language understanding tasks. We term this shortfall the text-speech understanding gap: the performance drop observed when a speech-adapted LLM processes spoken inputs relative to when the original text-based LLM processes the equivalent text. Recent approaches to narrowing this gap either rely on large-scale speech synthesis of text corpora, which is costly and heavily dependent on synthetic data, or on large-scale proprietary speech datasets, which are not reproducible. As a result, there remains a need for more data-efficient alternatives for closing the text-speech understanding gap. In this work, we analyze the gap as driven by two factors: (i) forgetting of text capabilities during adaptation, and (ii) cross-modal misalignment between speech and text. Based on this analysis, we introduce SALAD--Sample-efficient Alignment with Learning through Active selection and cross-modal Distillation--which combines cross-modal distillation with targeted synthetic data to improve alignment while mitigating forgetting. Applied to 3B and 7B LLMs, SALAD achieves competitive performance with a strong open-weight model across broad-domain benchmarks in knowledge, language understanding, and reasoning, while training on over an order of magnitude less speech data from public corpora.

☆ Towards Multimodal Query-Based Spatial Audio Source Extraction ICASSP 2026

Chenxin Yu, Hao Ma, Xu Li, Xiao-Lei Zhang, Mingjie Shao, Chi Zhang, Xuelong Li

Query-based audio source extraction seeks to recover a target source from a mixture conditioned on a query. Existing approaches are largely confined to single-channel audio, leaving the spatial information in multi-channel recordings underexploited. We introduce a query-based spatial audio source extraction framework for recovering dry target signals from first-order ambisonics (FOA) mixtures. Our method accepts either an audio prompt or a text prompt as condition input, enabling flexible end-to-end extraction. The core of our proposed model lies in a tri-axial Transformer that jointly models temporal, frequency, and spatial channel dependencies. The model uses contrastive language-audio pretraining (CLAP) embeddings to enable unified audio-text conditioning via feature-wise linear modulation (FiLM). To eliminate costly annotations and improve generalization, we propose a label-free data pipeline that dynamically generates spatial mixtures and corresponding targets for training. The result of our experiment with high separation quality demonstrates the efficacy of multimodal conditioning and tri-axial modeling. This work establishes a new paradigm for high-fidelity spatial audio separation in immersive applications.

comment: Submitted to ICASSP 2026

☆ Two Heads Are Better Than One: Audio-Visual Speech Error Correction with Dual Hypotheses

Sungnyun Kim, Kangwook Jang, Sungwoo Cho, Joon Son Chung, Hoirin Kim, Se-Young Yun

This paper introduces a new paradigm for generative error correction (GER) framework in audio-visual speech recognition (AVSR) that reasons over modality-specific evidences directly in the language space. Our framework, DualHyp, empowers a large language model (LLM) to compose independent N-best hypotheses from separate automatic speech recognition (ASR) and visual speech recognition (VSR) models. To maximize the effectiveness of DualHyp, we further introduce RelPrompt, a noise-aware guidance mechanism that provides modality-grounded prompts to the LLM. RelPrompt offers the temporal reliability of each modality stream, guiding the model to dynamically switch its focus between ASR and VSR hypotheses for an accurate correction. Under various corruption scenarios, our framework attains up to 57.7% error rate gain on the LRS2 benchmark over standard ASR baseline, contrary to single-stream GER approaches that achieve only 10% gain. To facilitate research within our DualHyp framework, we release the code and the dataset comprising ASR and VSR hypotheses at https://github.com/sungnyun/dualhyp.

comment: Preprint work

☆ Acoustic Teleportation via Disentangled Neural Audio Codec Representations

Philipp Grundhuber, Mhd Modar Halimeh, Emanuël A. P. Habets

This paper presents an approach for acoustic teleportation by disentangling speech content from acoustic environment characteristics in neural audio codec representations. Acoustic teleportation transfers room characteristics between speech recordings while preserving content and speaker identity. We build upon previous work using the EnCodec architecture, achieving substantial objective quality improvements with non-intrusive ScoreQ scores of 3.03, compared to 2.44 for prior methods. Our training strategy incorporates five tasks: clean reconstruction, reverberated reconstruction, dereverberation, and two variants of acoustic teleportation. We demonstrate that temporal downsampling of the acoustic embedding significantly degrades performance, with even 2x downsampling resulting in a statistically significant reduction in quality. The learned acoustic embeddings exhibit strong correlations with RT60. Effective disentanglement is demonstrated using t-SNE clustering analysis, where acoustic embeddings cluster by room while speech embeddings cluster by speaker.

♻ ☆ FakeMark: Deepfake Speech Attribution With Watermarked Artifacts

Wanying Ge, Xin Wang, Junichi Yamagishi

Deepfake speech attribution remains challenging for existing solutions. Classifier-based solutions often fail to generalize to domain-shifted samples, and watermarking-based solutions are easily compromised by distortions like codec compression or malicious removal attacks. To address these issues, we propose FakeMark, a novel watermarking framework that injects artifact-correlated watermarks associated with deepfake systems rather than pre-assigned bitstring messages. This design allows a detector to attribute the source system by leveraging both injected watermark and intrinsic deepfake artifacts, remaining effective even if one of these cues is elusive or removed. Experimental results show that FakeMark improves generalization to cross-dataset samples where classifier-based solutions struggle and maintains high accuracy under various distortions where conventional watermarking-based solutions fail.

♻ ☆ DiSTAR: Diffusion over a Scalable Token Autoregressive Representation for Speech Generation

Yakun Song, Xiaobin Zhuang, Jiawei Chen, Zhikang Niu, Guanrou Yang, Chenpeng Du, Dongya Jia, Zhuo Chen, Yuping Wang, Yuxuan Wang, Xie Chen

Recent attempts to interleave autoregressive (AR) sketchers with diffusion-based refiners over continuous speech representations have shown promise, but they remain brittle under distribution shift and offer limited levers for controllability. We introduce DISTAR, a zero-shot text-to-speech framework that operates entirely in a discrete residual vector quantization (RVQ) code space and tightly couples an AR language model with a masked diffusion model, without forced alignment or a duration predictor. Concretely, DISTAR drafts block-level RVQ tokens with an AR language model and then performs parallel masked-diffusion infilling conditioned on the draft to complete the next block, yielding long-form synthesis with blockwise parallelism while mitigating classic AR exposure bias. The discrete code space affords explicit control at inference: DISTAR produces high-quality audio under both greedy and sample-based decoding using classifier-free guidance, supports trade-offs between robustness and diversity, and enables variable bit-rate and controllable computation via RVQ layer pruning at test time. Extensive experiments and ablations demonstrate that DISTAR surpasses state-of-the-art zero-shot TTS systems in robustness, naturalness, and speaker/style consistency, while maintaining rich output diversity. Audio samples are provided on https://anonymous.4open.science/w/DiSTAR_demo.

♻ ☆ AudioGenie-Reasoner: A Training-Free Multi-Agent Framework for Coarse-to-Fine Audio Deep Reasoning

Yan Rong, Chenxing Li, Dong Yu, Li Liu

Audio deep reasoning is a challenging task that requires expert-level perception, multi-step logical inference, and the integration of contextual knowledge. However, existing models suffer from a gap between audio perception and reasoning abilities due to the lack of training data with explicit reasoning chains and the absence of mechanisms for active exploration and iterative refinement. To address these challenges, we propose AudioGenie-Reasoner (AGR), the first unified training-free multi-agent system that coordinates perception and reasoning over an evolving chain of textual evidence. Our key idea is a paradigm shift that transforms audio deep reasoning into complex text understanding task from a new perspective, thereby unlocking the full potential of large language models. Specifically, the design of AGR mimics the human coarse-to-fine cognitive process. It first transforms the input audio into a coarse text-based document. Then, we design a novel proactive iterative document refinement loop, featuring tool-augmented routes and specialized agents, to continuously search for missing information and augment the evidence chain in a coarse-to-fine manner until sufficient question-related information is gathered for making final predictions. Experimental results show that AGR achieves state-of-the-art (SOTA) performance over existing open-source audio deep reasoning models across various benchmarks. The code will be available at https://github.com/ryysayhi/AudioGenie-Reasoner.

♻ ☆ MelCap: A Unified Single-Codebook Neural Codec for High-Fidelity Audio Compression

Jingyi Li, Zhiyuan Zhao, Yunfei Liu, Lijian Lin, Ye Zhu, Jiahao Wu, Qiuqiang Kong, Yu Li

Neural audio codecs have recently emerged as powerful tools for high-quality and low-bitrate audio compression, leveraging deep generative models to learn latent representations of audio signals. However, existing approaches either rely on a single quantizer that only processes speech domain, or on multiple quantizers that are not well suited for downstream tasks. To address this issue, we propose MelCap, a unified "one-codebook-for-all" neural codec that effectively handles speech, music, and general sound. By decomposing audio reconstruction into two stages, our method preserves more acoustic details than previous single-codebook approaches, while achieving performance comparable to mainstream multi-codebook methods. In the first stage, audio is transformed into mel-spectrograms, which are compressed and quantized into compact single tokens using a 2D tokenizer. A perceptual loss is further applied to mitigate the over-smoothing artifacts observed in spectrogram reconstruction. In the second stage, a Vocoder recovers waveforms from the mel discrete tokens in a single forward pass, enabling real-time decoding. Both objective and subjective evaluations demonstrate that MelCap achieves quality on comparable to state-of-the-art multi-codebook codecs, while retaining the computational simplicity of a single-codebook design, thereby providing an effective representation for downstream tasks.

comment: 9 pages, 4 figures

♻ ☆ ASE: Practical Acoustic Speed Estimation Beyond Doppler via Sound Diffusion Field

Sheng Lyu, Chenshu Wu

Passive human speed estimation plays a critical role in acoustic sensing. Despite extensive study, existing systems, however, suffer from various limitations: First, the channel measurement rate proves inadequate to estimate high moving speeds. Second, previous acoustic speed estimation exploits Doppler Frequency Shifts (DFS) created by moving targets and relies on microphone arrays, making them only capable of sensing the radial speed within a constrained distance. To overcome these issues, we present ASE, an accurate and robust Acoustic Speed Estimation system on a single commodity microphone. We propose a novel Orthogonal Time-Delayed Multiplexing (OTDM) scheme for acoustic channel estimation at a high rate that was previously infeasible, making it possible to estimate high speeds. We then model the sound propagation from a unique perspective of the acoustic diffusion field, and infer the speed from the acoustic spatial distribution, a completely different way of thinking about speed estimation beyond prior DFS-based approaches. We further develop novel techniques for motion detection and signal enhancement to deliver a robust and practical system. We implement and evaluate ASE through extensive real-world experiments. Our results show that ASE reliably tracks walking speed, independently of target location and direction, with a mean error of 0.13 m/s, a reduction of 2.5x from DFS, and a detection rate of 97.4% for large coverage, e.g., free walking in a 4m x 4m room. We believe ASE pushes acoustic speed estimation beyond the conventional DFS-based paradigm and inspires exciting research in acoustic sensing. Code is available at https://github.com/aiot-lab/ASE.

comment: Accepted at IMWUT'25

♻ ☆ Universal Speech Token Learning via Low-Bitrate Neural Codec and Pretrained Representations SP

Xue Jiang, Xiulian Peng, Yuan Zhang, Yan Lu

Current large speech language models are mainly based on semantic tokens from discretization of self-supervised learned representations and acoustic tokens from a neural codec, following a semantic-modeling and acoustic-synthesis paradigm. However, semantic tokens discard paralinguistic attributes of speakers that is important for natural spoken communication, while prompt-based acoustic synthesis from semantic tokens has limits in recovering paralinguistic details and suffers from robustness issues, especially when there are domain gaps between the prompt and the target. This paper unifies two types of tokens and proposes the UniCodec, a universal speech token learning that encapsulates all semantics of speech, including linguistic and paralinguistic information, into a compact and semantically-disentangled unified token. Such a unified token can not only benefit speech language models in understanding with paralinguistic hints but also help speech generation with high-quality output. A low-bitrate neural codec is leveraged to learn such disentangled discrete representations at global and local scales, with knowledge distilled from self-supervised learned features. Extensive evaluations on multilingual datasets demonstrate its effectiveness in generating natural, expressive and long-term consistent output quality with paralinguistic attributes well preserved in several speech processing tasks.

comment: Accepted by IEEE Journal of Selected Topics in Signal Processing(JSTSP)

♻ ☆ Latent-Domain Predictive Neural Speech Coding SP

Xue Jiang, Xiulian Peng, Huaying Xue, Yuan Zhang, Yan Lu

Neural audio/speech coding has recently demonstrated its capability to deliver high quality at much lower bitrates than traditional methods. However, existing neural audio/speech codecs employ either acoustic features or learned blind features with a convolutional neural network for encoding, by which there are still temporal redundancies within encoded features. This paper introduces latent-domain predictive coding into the VQ-VAE framework to fully remove such redundancies and proposes the TF-Codec for low-latency neural speech coding in an end-to-end manner. Specifically, the extracted features are encoded conditioned on a prediction from past quantized latent frames so that temporal correlations are further removed. Moreover, we introduce a learnable compression on the time-frequency input to adaptively adjust the attention paid to main frequencies and details at different bitrates. A differentiable vector quantization scheme based on distance-to-soft mapping and Gumbel-Softmax is proposed to better model the latent distributions with rate constraint. Subjective results on multilingual speech datasets show that, with low latency, the proposed TF-Codec at 1 kbps achieves significantly better quality than Opus at 9 kbps, and TF-Codec at 3 kbps outperforms both EVS at 9.6 kbps and Opus at 12 kbps. Numerous studies are conducted to demonstrate the effectiveness of these techniques. Code and models are available at https://github.com/microsoft/TF-Codec.

comment: Accepted by IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING (TASLP). Code and models are available at https://github.com/microsoft/TF-Codec

♻ ☆ MSR-Codec: A Low-Bitrate Multi-Stream Residual Codec for High-Fidelity Speech Generation with Information Disentanglement

Jingyu Li, Guangyan Zhang, Zhen Ye, Yiwen Guo

Audio codecs are a critical component of modern speech generation systems. This paper introduces a low-bitrate, multi-scale residual codec that encodes speech into four distinct streams: semantic, timbre, prosody, and residual. This architecture achieves high-fidelity speech reconstruction at competitive low bitrates while demonstrating an inherent ability for information disentanglement. We construct a two-stage language model for text-to-speech (TTS) synthesis using this codec, which, despite its lightweight design and minimal data requirements, achieves a state-of-the-art Word Error Rate (WER) and superior speaker similarity compared to several larger models. Furthermore, the codec's design proves highly effective for voice conversion, enabling independent manipulation of speaker timbre and prosody. Our inference code, pre-trained models, and audio samples are available at https://github.com/herbertLJY/MSRCodec.

♻ ☆ SAGE-Music: Low-Latency Symbolic Music Generation via Attribute-Specialized Key-Value Head Sharing

Jiaye Tan, Haonan Luo, Linfeng Song, Shuaiqi Chen, Yishan Lyu, Zian Zhong, Roujia Wang, Daniel Jiang, Haoran Zhang, Jiaming Bai, Haoran Cheng, Q. Vera Liao, Hao-Wen Dong

Low-latency symbolic music generation is essential for real-time improvisation and human-AI co-creation. Existing transformer-based models, however, face a trade-off between inference speed and musical quality. Traditional acceleration techniques such as embedding pooling significantly degrade quality, while recently proposed Byte Pair Encoding (BPE) methods - though effective on single-track piano data - suffer large performance drops in multi-track settings, as revealed by our analysis. We propose Attribute-Specialized Key-Value Head Sharing (AS-KVHS), adapted to music's structured symbolic representation, achieving about 30% inference speedup with only a negligible (about 0.4%) quality drop in objective evaluations and slight improvements in subjective listening tests. Our main contributions are (1) the first systematic study of BPE's generalizability in multi-track symbolic music, and (2) the introduction of AS-KVHS for low-latency symbolic music generation. Beyond these, we also release SAGE-Music, an open-source benchmark that matches or surpasses state-of-the-art models in generation quality.

comment: Withdrawn after identifying that results in Section 5 require additional re-analysis before public dissemination

☆ Musical consonance: a review of theory and evidence on perception and preference of auditory roughness in humans and other animals

John M. McBride

The origins of consonance in human music has long been contested, and today there are three primary hypotheses: aversion to roughness, preference for harmonicity, and learned preferences from cultural exposure. While the evidence is currently insufficient to disentangle the contributions of these hypotheses, I propose several reasons why roughness is an especially promising area for future study. The aim of this review is to summarize and critically evaluate roughness theory and models, experimental data, to highlight areas that deserve further research. I identify 2 key areas: There are fundamental issues with the definition and interpretation of results due to tautology in the definition of roughness, and the lack of independence in empirical measurements. Despite extensive model development, there are many duplications and models have issues with data quality and overfitting. Future theory development should aim for model simplicity, and extra assumptions, features and parameters should be evaluated systematically. Model evaluation should aim to maximise the breadth of stimuli that are predicted.

☆ Steer-MoE: Efficient Audio-Language Alignment with a Mixture-of-Experts Steering Module ICASSP 2026

Ruitao Feng, Bixi Zhang, Sheng Liang, Zheng Yuan

Aligning pretrained audio encoders and Large Language Models (LLMs) offers a promising, parameter-efficient path to building powerful multimodal agents. However, existing methods often require costly full-model finetuning or rely on static adapters that may lack expressive power. Drawing inspiration from the Platonic Representation Hypothesis, we introduce SteerMoE, a novel and modular framework for audio-language alignment. SteerMoE freezes both the audio encoder and the LLM decoder, training only a lightweight steering module integrated within the encoder's layers. This module uses a Mixture-of-Experts (MoE) router to dynamically select and apply learned steering vectors, progressively transforming continuous audio representations into a space comprehensible to the LLM. By operating entirely in the continuous embedding space, our approach requires no modifications to the LLM's vocabulary and preserves its advanced reasoning and agentic capabilities. We demonstrate through experiments on ASR, audio understanding, and a qualitative function-calling task that SteerMoE achieves strong performance while remaining highly modular and computationally efficient, offering a robust new paradigm for developing sophisticated audio-language systems.

comment: 5 pages, 1 figures. Code is available at: https://github.com/forfrt/SteerMoE. Submitted to ICASSP 2026

☆ UniMoE-Audio: Unified Speech and Music Generation with Dynamic-Capacity MoE

Zhenyu Liu, Yunxin Li, Xuanyu Zhang, Qixun Teng, Shenyuan Jiang, Xinyu Chen, Haoyuan Shi, Jinchao Li, Qi Wang, Haolan Chen, Fanbo Meng, Mingjun Zhao, Yu Xu, Yancheng He, Baotian Hu, Min Zhang

Recent advances in unified multimodal models indicate a clear trend towards comprehensive content generation. However, the auditory domain remains a significant challenge, with music and speech often developed in isolation, hindering progress towards universal audio synthesis. This separation stems from inherent task conflicts and severe data imbalances, which impede the development of a truly unified audio generation model. To address this challenge, we propose UniMoE-Audio, a unified speech and music generation model within a novel Dynamic-Capacity Mixture-of-Experts (MoE) framework. Architecturally, UniMoE-Audio introduces a Top-P routing strategy for dynamic expert number allocation, and a hybrid expert design comprising routed experts for domain-specific knowledge, shared experts for domain-agnostic features, and null experts for adaptive computation skipping. To tackle data imbalance, we introduce a three-stage training curriculum: 1) Independent Specialist Training leverages original datasets to instill domain-specific knowledge into each "proto-expert" without interference; 2) MoE Integration and Warmup incorporates these specialists into the UniMoE-Audio architecture, warming up the gate module and shared expert using a subset of balanced dataset; and 3) Synergistic Joint Training trains the entire model end-to-end on the fully balanced dataset, fostering enhanced cross-domain synergy. Extensive experiments show that UniMoE-Audio not only achieves state-of-the-art performance on major speech and music generation benchmarks, but also demonstrates superior synergistic learning, mitigating the performance degradation typically seen in naive joint training. Our findings highlight the substantial potential of specialized MoE architecture and curated training strategies in advancing the field of universal audio generation. Homepage: https://mukioxun.github.io/Uni-MoE-site/home.html

☆ MotionBeat: Motion-Aligned Music Representation via Embodied Contrastive Learning and Bar-Equivariant Contact-Aware Encoding

Xuanchen Wang, Heng Wang, Weidong Cai

Music is both an auditory and an embodied phenomenon, closely linked to human motion and naturally expressed through dance. However, most existing audio representations neglect this embodied dimension, limiting their ability to capture rhythmic and structural cues that drive movement. We propose MotionBeat, a framework for motion-aligned music representation learning. MotionBeat is trained with two newly proposed objectives: the Embodied Contrastive Loss (ECL), an enhanced InfoNCE formulation with tempo-aware and beat-jitter negatives to achieve fine-grained rhythmic discrimination, and the Structural Rhythm Alignment Loss (SRAL), which ensures rhythm consistency by aligning music accents with corresponding motion events. Architecturally, MotionBeat introduces bar-equivariant phase rotations to capture cyclic rhythmic patterns and contact-guided attention to emphasize motion events synchronized with musical accents. Experiments show that MotionBeat outperforms state-of-the-art audio encoders in music-to-dance generation and transfers effectively to beat tracking, music tagging, genre and instrument classification, emotion recognition, and audio-visual retrieval. Our project demo page: https://motionbeat2025.github.io/.

comment: 5 pages, 1 figure. demo page: https://motionbeat2025.github.io/

♻ ☆ AudioGenie-Reasoner: A Training-Free Multi-Agent Framework for Coarse-to-Fine Audio Deep Reasoning

Yan Rong, Chenxing Li, Dong Yu, Li Liu

Audio deep reasoning is a challenging task that requires expert-level perception, multi-step logical inference, and the integration of contextual knowledge. However, existing models suffer from a gap between audio perception and reasoning abilities due to the lack of training data with explicit reasoning chains and the absence of mechanisms for active exploration and iterative refinement. To address these challenges, we propose AudioGenie-Reasoner (AGR), the first unified training-free multi-agent system that coordinates perception and reasoning over an evolving chain of textual evidence. Our key idea is a paradigm shift that transforms audio deep reasoning into complex text understanding task from a new perspective, thereby unlocking the full potential of large language models. Specifically, the design of AGR mimics the human coarse-to-fine cognitive process. It first transforms the input audio into a coarse text-based document. Then, we design a novel proactive iterative document refinement loop, featuring tool-augmented routes and specialized agents, to continuously search for missing information and augment the evidence chain in a coarse-to-fine manner until sufficient question-related information is gathered for making final predictions. Experimental results show that AGR achieves state-of-the-art (SOTA) performance over existing open-source audio deep reasoning models across various benchmarks. The code will be available at https://github.com/ryysayhi/AudioGenie-Reasoner.

♻ ☆ MelCap: A Unified Single-Codebook Neural Codec for High-Fidelity Audio Compression

Jingyi Li, Zhiyuan Zhao, Yunfei Liu, Lijian Lin, Ye Zhu, Jiahao Wu, Qiuqiang Kong, Yu Li

Neural audio codecs have recently emerged as powerful tools for high-quality and low-bitrate audio compression, leveraging deep generative models to learn latent representations of audio signals. However, existing approaches either rely on a single quantizer that only processes speech domain, or on multiple quantizers that are not well suited for downstream tasks. To address this issue, we propose MelCap, a unified "one-codebook-for-all" neural codec that effectively handles speech, music, and general sound. By decomposing audio reconstruction into two stages, our method preserves more acoustic details than previous single-codebook approaches, while achieving performance comparable to mainstream multi-codebook methods. In the first stage, audio is transformed into mel-spectrograms, which are compressed and quantized into compact single tokens using a 2D tokenizer. A perceptual loss is further applied to mitigate the over-smoothing artifacts observed in spectrogram reconstruction. In the second stage, a Vocoder recovers waveforms from the mel discrete tokens in a single forward pass, enabling real-time decoding. Both objective and subjective evaluations demonstrate that MelCap achieves quality on comparable to state-of-the-art multi-codebook codecs, while retaining the computational simplicity of a single-codebook design, thereby providing an effective representation for downstream tasks.

comment: 9 pages, 4 figures

♻ ☆ ASE: Practical Acoustic Speed Estimation Beyond Doppler via Sound Diffusion Field

Sheng Lyu, Chenshu Wu

Passive human speed estimation plays a critical role in acoustic sensing. Despite extensive study, existing systems, however, suffer from various limitations: First, the channel measurement rate proves inadequate to estimate high moving speeds. Second, previous acoustic speed estimation exploits Doppler Frequency Shifts (DFS) created by moving targets and relies on microphone arrays, making them only capable of sensing the radial speed within a constrained distance. To overcome these issues, we present ASE, an accurate and robust Acoustic Speed Estimation system on a single commodity microphone. We propose a novel Orthogonal Time-Delayed Multiplexing (OTDM) scheme for acoustic channel estimation at a high rate that was previously infeasible, making it possible to estimate high speeds. We then model the sound propagation from a unique perspective of the acoustic diffusion field, and infer the speed from the acoustic spatial distribution, a completely different way of thinking about speed estimation beyond prior DFS-based approaches. We further develop novel techniques for motion detection and signal enhancement to deliver a robust and practical system. We implement and evaluate ASE through extensive real-world experiments. Our results show that ASE reliably tracks walking speed, independently of target location and direction, with a mean error of 0.13 m/s, a reduction of 2.5x from DFS, and a detection rate of 97.4% for large coverage, e.g., free walking in a 4m x 4m room. We believe ASE pushes acoustic speed estimation beyond the conventional DFS-based paradigm and inspires exciting research in acoustic sensing. Code is available at https://github.com/aiot-lab/ASE.

comment: Accepted at IMWUT'25

♻ ☆ Universal Speech Token Learning via Low-Bitrate Neural Codec and Pretrained Representations SP

Xue Jiang, Xiulian Peng, Yuan Zhang, Yan Lu

Current large speech language models are mainly based on semantic tokens from discretization of self-supervised learned representations and acoustic tokens from a neural codec, following a semantic-modeling and acoustic-synthesis paradigm. However, semantic tokens discard paralinguistic attributes of speakers that is important for natural spoken communication, while prompt-based acoustic synthesis from semantic tokens has limits in recovering paralinguistic details and suffers from robustness issues, especially when there are domain gaps between the prompt and the target. This paper unifies two types of tokens and proposes the UniCodec, a universal speech token learning that encapsulates all semantics of speech, including linguistic and paralinguistic information, into a compact and semantically-disentangled unified token. Such a unified token can not only benefit speech language models in understanding with paralinguistic hints but also help speech generation with high-quality output. A low-bitrate neural codec is leveraged to learn such disentangled discrete representations at global and local scales, with knowledge distilled from self-supervised learned features. Extensive evaluations on multilingual datasets demonstrate its effectiveness in generating natural, expressive and long-term consistent output quality with paralinguistic attributes well preserved in several speech processing tasks.

comment: Accepted by IEEE Journal of Selected Topics in Signal Processing(JSTSP)

♻ ☆ Latent-Domain Predictive Neural Speech Coding SP

Xue Jiang, Xiulian Peng, Huaying Xue, Yuan Zhang, Yan Lu

Neural audio/speech coding has recently demonstrated its capability to deliver high quality at much lower bitrates than traditional methods. However, existing neural audio/speech codecs employ either acoustic features or learned blind features with a convolutional neural network for encoding, by which there are still temporal redundancies within encoded features. This paper introduces latent-domain predictive coding into the VQ-VAE framework to fully remove such redundancies and proposes the TF-Codec for low-latency neural speech coding in an end-to-end manner. Specifically, the extracted features are encoded conditioned on a prediction from past quantized latent frames so that temporal correlations are further removed. Moreover, we introduce a learnable compression on the time-frequency input to adaptively adjust the attention paid to main frequencies and details at different bitrates. A differentiable vector quantization scheme based on distance-to-soft mapping and Gumbel-Softmax is proposed to better model the latent distributions with rate constraint. Subjective results on multilingual speech datasets show that, with low latency, the proposed TF-Codec at 1 kbps achieves significantly better quality than Opus at 9 kbps, and TF-Codec at 3 kbps outperforms both EVS at 9.6 kbps and Opus at 12 kbps. Numerous studies are conducted to demonstrate the effectiveness of these techniques. Code and models are available at https://github.com/microsoft/TF-Codec.

comment: Accepted by IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING (TASLP). Code and models are available at https://github.com/microsoft/TF-Codec

♻ ☆ SAGE-Music: Low-Latency Symbolic Music Generation via Attribute-Specialized Key-Value Head Sharing

Jiaye Tan, Haonan Luo, Linfeng Song, Shuaiqi Chen, Yishan Lyu, Zian Zhong, Roujia Wang, Daniel Jiang, Haoran Zhang, Jiaming Bai, Haoran Cheng, Q. Vera Liao, Hao-Wen Dong

Low-latency symbolic music generation is essential for real-time improvisation and human-AI co-creation. Existing transformer-based models, however, face a trade-off between inference speed and musical quality. Traditional acceleration techniques such as embedding pooling significantly degrade quality, while recently proposed Byte Pair Encoding (BPE) methods - though effective on single-track piano data - suffer large performance drops in multi-track settings, as revealed by our analysis. We propose Attribute-Specialized Key-Value Head Sharing (AS-KVHS), adapted to music's structured symbolic representation, achieving about 30% inference speedup with only a negligible (about 0.4%) quality drop in objective evaluations and slight improvements in subjective listening tests. Our main contributions are (1) the first systematic study of BPE's generalizability in multi-track symbolic music, and (2) the introduction of AS-KVHS for low-latency symbolic music generation. Beyond these, we also release SAGE-Music, an open-source benchmark that matches or surpasses state-of-the-art models in generation quality.

comment: Withdrawn after identifying that results in Section 5 require additional re-analysis before public dissemination

☆ Content Anonymization for Privacy in Long-form Audio

Cristina Aggazzotti, Ashi Garg, Zexin Cai, Nicholas Andrews

Voice anonymization techniques have been found to successfully obscure a speaker's acoustic identity in short, isolated utterances in benchmarks such as the VoicePrivacy Challenge. In practice, however, utterances seldom occur in isolation: long-form audio is commonplace in domains such as interviews, phone calls, and meetings. In these cases, many utterances from the same speaker are available, which pose a significantly greater privacy risk: given multiple utterances from the same speaker, an attacker could exploit an individual's vocabulary, syntax, and turns of phrase to re-identify them, even when their voice is completely disguised. To address this risk, we propose new content anonymization approaches. Our approach performs a contextual rewriting of the transcripts in an ASR-TTS pipeline to eliminate speaker-specific style while preserving meaning. We present results in a long-form telephone conversation setting demonstrating the effectiveness of a content-based attack on voice-anonymized speech. Then we show how the proposed content-based anonymization methods can mitigate this risk while preserving speech utility. Overall, we find that paraphrasing is an effective defense against content-based attacks and recommend that stakeholders adopt this step to ensure anonymity in long-form audio.

☆ Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception

Ziyang Ma, Ruiyang Xu, Zhenghao Xing, Yunfei Chu, Yuxuan Wang, Jinzheng He, Jin Xu, Pheng-Ann Heng, Kai Yu, Junyang Lin, Eng Siong Chng, Xie Chen

Fine-grained perception of multimodal information is critical for advancing human-AI interaction. With recent progress in audio-visual technologies, Omni Language Models (OLMs), capable of processing audio and video signals in parallel, have emerged as a promising paradigm for achieving richer understanding and reasoning. However, their capacity to capture and describe fine-grained details remains limited explored. In this work, we present a systematic and comprehensive investigation of omni detailed perception from the perspectives of the data pipeline, models, and benchmark. We first identify an inherent "co-growth" between detail and hallucination in current OLMs. To address this, we propose Omni-Detective, an agentic data generation pipeline integrating tool-calling, to autonomously produce highly detailed yet minimally hallucinatory multimodal data. Based on the data generated with Omni-Detective, we train two captioning models: Audio-Captioner for audio-only detailed perception, and Omni-Captioner for audio-visual detailed perception. Under the cascade evaluation protocol, Audio-Captioner achieves the best performance on MMAU and MMAR among all open-source models, surpassing Gemini 2.5 Flash and delivering performance comparable to Gemini 2.5 Pro. On existing detailed captioning benchmarks, Omni-Captioner sets a new state-of-the-art on VDC and achieves the best trade-off between detail and hallucination on the video-SALMONN 2 testset. Given the absence of a dedicated benchmark for omni detailed perception, we design Omni-Cloze, a novel cloze-style evaluation for detailed audio, visual, and audio-visual captioning that ensures stable, efficient, and reliable assessment. Experimental results and analysis demonstrate the effectiveness of Omni-Detective in generating high-quality detailed captions, as well as the superiority of Omni-Cloze in evaluating such detailed captions.

comment: https://github.com/ddlBoJack/Omni-Captioner

☆ TFGA-Net: Temporal-Frequency Graph Attention Network for Brain-Controlled Speaker Extraction

Youhao Si, Yuan Liao, Qiushi Han, Yuhang Yang, Rui Dai, Liya Huang

The rapid development of auditory attention decoding (AAD) based on electroencephalography (EEG) signals offers the possibility EEG-driven target speaker extraction. However, how to effectively utilize the target-speaker common information between EEG and speech remains an unresolved problem. In this paper, we propose a model for brain-controlled speaker extraction, which utilizes the EEG recorded from the listener to extract the target speech. In order to effectively extract information from EEG signals, we derive multi-scale time--frequency features and further incorporate cortical topological structures that are selectively engaged during the task. Moreover, to effectively exploit the non-Euclidean structure of EEG signals and capture their global features, the graph convolutional networks and self-attention mechanism are used in the EEG encoder. In addition, to make full use of the fused EEG and speech feature and preserve global context and capture speech rhythm and prosody, we introduce MossFormer2 which combines MossFormer and RNN-Free Recurrent as separator. Experimental results on both the public Cocktail Party and KUL dataset in this paper show that our TFGA-Net model significantly outper-forms the state-of-the-art method in certain objective evaluation metrics. The source code is available at: https://github.com/LaoDa-X/TFGA-NET.

comment: 5 pages, 3 figures

☆ Not in Sync: Unveiling Temporal Bias in Audio Chat Models

Jiayu Yao, Shenghua Liu, Yiwei Wang, Rundong Cheng, Lingrui Mei, Baolong Bi, Zhen Xiong, Xueqi Cheng

Large Audio Language Models (LALMs) are increasingly applied to audio understanding and multimodal reasoning, yet their ability to locate when events occur remains underexplored. We present the first systematic study of temporal bias in LALMs, revealing a key limitation in their timestamp prediction. For example, when asked "At which second does the lecturer introduce the key formula?", models often predict timestamps that are consistently earlier or later than the ground truth. Through controlled experiments on timestamped datasets, we find that temporal bias (i) is prevalent across datasets and models, (ii) increases with audio length - even accumulating to tens of seconds in extended recordings, and (iii) varies across event types and positions. We quantify this effect with the Temporal Bias Index (TBI), measuring systematic misalignment in predicted event timings, and complement it with a visualization framework. Our findings highlight a fundamental limitation in current LALMs and call for the development of temporally robust architectures.

☆ Audio Palette: A Diffusion Transformer with Multi-Signal Conditioning for Controllable Foley Synthesis

Junnuo Wang

Recent advances in diffusion-based generative models have enabled high-quality text-to-audio synthesis, but fine-grained acoustic control remains a significant challenge in open-source research. We present Audio Palette, a diffusion transformer (DiT) based model that extends the Stable Audio Open architecture to address this "control gap" in controllable audio generation. Unlike prior approaches that rely solely on semantic conditioning, Audio Palette introduces four time-varying control signals: loudness, pitch, spectral centroid, and timbre, for precise and interpretable manipulation of acoustic features. The model is efficiently adapted for the nuanced domain of Foley synthesis using Low-Rank Adaptation (LoRA) on a curated subset of AudioSet, requiring only 0.85 percent of the original parameters to be trained. Experiments demonstrate that Audio Palette achieves fine-grained, interpretable control of sound attributes. Crucially, it accomplishes this novel controllability while maintaining high audio quality and strong semantic alignment to text prompts, with performance on standard metrics such as Frechet Audio Distance (FAD) and LAION-CLAP scores remaining comparable to the original baseline model. We provide a scalable, modular pipeline for audio research, emphasizing sequence-based conditioning, memory efficiency, and a three-scale classifier-free guidance mechanism for nuanced inference-time control. This work establishes a robust foundation for controllable sound design and performative audio synthesis in open-source settings, enabling a more artist-centric workflow.

comment: Accepted for publication in the Journal of Artificial Intelligence Research (JAIR), Vol. 3 No. 2, December 2025

☆ Switchboard-Affect: Emotion Perception Labels from Conversational Speech

Amrit Romana, Jaya Narain, Tien Dung Tran, Andrea Davis, Jason Fong, Ramya Rasipuram, Vikramjit Mitra

Understanding the nuances of speech emotion dataset curation and labeling is essential for assessing speech emotion recognition (SER) model potential in real-world applications. Most training and evaluation datasets contain acted or pseudo-acted speech (e.g., podcast speech) in which emotion expressions may be exaggerated or otherwise intentionally modified. Furthermore, datasets labeled based on crowd perception often lack transparency regarding the guidelines given to annotators. These factors make it difficult to understand model performance and pinpoint necessary areas for improvement. To address this gap, we identified the Switchboard corpus as a promising source of naturalistic conversational speech, and we trained a crowd to label the dataset for categorical emotions (anger, contempt, disgust, fear, sadness, surprise, happiness, tenderness, calmness, and neutral) and dimensional attributes (activation, valence, and dominance). We refer to this label set as Switchboard-Affect (SWB-Affect). In this work, we present our approach in detail, including the definitions provided to annotators and an analysis of the lexical and paralinguistic cues that may have played a role in their perception. In addition, we evaluate state-of-the-art SER models, and we find variable performance across the emotion categories with especially poor generalization for anger. These findings underscore the importance of evaluation with datasets that capture natural affective variations in speech. We release the labels for SWB-Affect to enable further analysis in this domain.

comment: 2025 13th International Conference on Affective Computing and Intelligent Interaction (ACII) https://github.com/apple/ml-switchboard-affect

☆ Continuous-Token Diffusion for Speaker-Referenced TTS in Multimodal LLMs

Xinlu He, Swayambhu Nath Ray, Harish Mallidi, Jia-Hong Huang, Ashwin Bellur, Chander Chandak, M. Maruf, Venkatesh Ravichandran

Unified architectures in multimodal large language models (MLLM) have shown promise in handling diverse tasks within a single framework. In the text-to-speech (TTS) task, current MLLM-based approaches rely on discrete token representations, which disregard the inherently continuous nature of speech and can lead to loss of fine-grained acoustic information.In this work, we investigate the TTS within the MLLM paradigm using continuous speech representations. We design a dual-head architecture and implement two complementary training strategies for a robust model. (1) A diffusion head generating continuous speech representations is added on the MLLM, which is on frame-level and strictly autoregressive. (2) The original language model head is retained to preserve multitask capability and to control the start and end of speech synthesis. (3) Masked training is employed to address exposure bias in autoregressive decoding. (4) To stabilize optimization, we propose a two-stage scheme where the LM is frozen in the second stage, ensuring the diffusion head learns from a fixed input distribution. Evaluations on LibriSpeech(PC) test-clean show that our approach achieves state-of-the-art autoregressive performance, with a WER of 1.95%, speaker similarity of 0.54, and UTMOS of 4.00. The two-stage training yields a 46% relative WER reduction over the one-stage training baseline. These results highlight the effectiveness of combining autoregressive modeling with continuous-token diffusion, supported by a two-stage training procedure.

☆ VCTR: A Transformer-Based Model for Non-parallel Voice Conversion

Maharnab Saikia

Non-parallel voice conversion aims to convert voice from a source domain to a target domain without paired training data. Cycle-Consistent Generative Adversarial Networks (CycleGAN) and Variational Autoencoders (VAE) have been used for this task, but these models suffer from difficult training and unsatisfactory results. Later, Contrastive Voice Conversion (CVC) was introduced, utilizing a contrastive learning-based approach to address these issues. However, these methods use CNN-based generators, which can capture local semantics but lacks the ability to capture long-range dependencies necessary for global semantics. In this paper, we propose VCTR, an efficient method for non-parallel voice conversion that leverages the Hybrid Perception Block (HPB) and Dual Pruned Self-Attention (DPSA) along with a contrastive learning-based adversarial approach. The code can be found in https://github.com/Maharnab-Saikia/VCTR.

☆ HyWA: Hypernetwork Weight Adapting Personalized Voice Activity Detection

Mahsa Ghazvini Nejad, Hamed Jafarzadeh Asl, Amin Edraki, Mohammadreza Sadeghi, Masoud Asgharian, Yuanhao Yu, Vahid Partovi Nia

Personalized Voice Activity Detection (PVAD) systems activate only in response to a specific target speaker by incorporating speaker embeddings from enrollment utterances. Unlike existing methods that require architectural changes, such as FiLM layers, our approach employs a hypernetwork to modify the weights of a few selected layers within a standard voice activity detection (VAD) model. This enables speaker conditioning without changing the VAD architecture, allowing the same VAD model to adapt to different speakers by updating only a small subset of the layers. We propose HyWA-PVAD, a hypernetwork weight adaptation method, and evaluate it against multiple baseline conditioning techniques. Our comparison shows consistent improvements in PVAD performance. HyWA also offers practical advantages for deployment by preserving the core VAD architecture. Our new approach improves the current conditioning techniques in two ways: i) increases the mean average precision, ii) simplifies deployment by reusing the same VAD architecture.

comment: Mahsa Ghazvini Nejad and Hamed Jafarzadeh Asl contributed equally to this work

☆ A Critical Review of the Need for Knowledge-Centric Evaluation of Quranic Recitation

Mohammed Hilal Al-Kharusi, Khizar Hayat, Khalil Bader Al Ruqeishi, Haroon Rashid Lone

The sacred practice of Quranic recitation (Tajweed), governed by precise phonetic, prosodic, and theological rules, faces significant pedagogical challenges in the modern era. While digital technologies promise unprecedented access to education, automated tools for recitation evaluation have failed to achieve widespread adoption or pedagogical efficacy. This literature review investigates this critical gap, conducting a comprehensive analysis of academic research, web platforms, and commercial applications developed over the past two decades. Our synthesis reveals a fundamental misalignment in prevailing approaches that repurpose Automatic Speech Recognition (ASR) architectures, which prioritize lexical recognition over qualitative acoustic assessment and are plagued by data dependency, demographic biases, and an inability to provide diagnostically useful feedback. Critiquing these data--driven paradigms, we argue for a foundational paradigm shift towards a knowledge-centric computational framework. Capitalizing on the immutable nature of the Quranic text and the precisely defined rules of Tajweed, we propose that a robust evaluator must be architected around anticipatory acoustic modeling based on canonical rules and articulation points (Makhraj), rather than relying on statistical patterns learned from imperfect and biased datasets. This review concludes that the future of automated Quranic evaluation lies in hybrid systems that integrate deep linguistic knowledge with advanced audio analysis, offering a path toward robust, equitable, and pedagogically sound tools that can faithfully support learners worldwide.

comment: 33 pages

☆ Adaptive vector steering: A training-free, layer-wise intervention for hallucination mitigation in large audio and multimodal models ICASSP 2026

Tsung-En Lin, Kuan-Yi Lee, Hung-Yi Lee

Large Audio-Language Models and Multi-Modal Large Language Models have demonstrated strong capabilities in tasks such as Audio Question Answering (AQA), Audio Captioning, and Automatic Speech Recognition (ASR). However, there is growing evidence that these models can hallucinate about the content of the audio. To address this issue, we probe the models' internal states and propose Adaptive Vector Steering (AVS), a method that better grounds generation in audio content. We also identify a strong correlation between output correctness and internal representations. Experiments show consistent performance gains across two models and two benchmarks. On the Audio Hallucination QA dataset, our method boosts the F1-score of Gemma from 0.550 to 0.619 and Qwen from 0.626 to 0.632. Furthermore, our method increases the accuracy of Qwen on MMAU from 0.548 to 0.592, marking an 8% relative increase. To the best of our knowledge, this is the first work to apply vector steering to mitigate hallucination in audio.

comment: Note: This preprint is a version of the paper submitted to ICASSP 2026. The author list here includes contributors who provided additional supervision and guidance. The official ICASSP submission may differ slightly in author composition

♻ ☆ Assessing Latency in ASR Systems: A Methodological Perspective for Real-Time Use

Carlos Arriaga, Alejandro Pozo, Javier Conde, Alvaro Alonso

Automatic speech recognition (ASR) systems generate real-time transcriptions but often miss nuances that human interpreters capture. While ASR is useful in many contexts, interpreters-who already use ASR tools such as Dragon-add critical value, especially in sensitive settings such as diplomatic meetings where subtle language is key. Human interpreters not only perceive these nuances but can adjust in real time, improving accuracy, while ASR handles basic transcription tasks. However, ASR systems introduce a delay that does not align with real-time interpretation needs. The user-perceived latency of ASR systems differs from that of interpretation because it measures the time between speech and transcription delivery. To address this, we propose a new approach to measuring delay in ASR systems and validate if they are usable in live interpretation scenarios.

comment: 8 pages, 2 figures

♻ ☆ Padé Approximant Neural Networks for Enhanced Electric Motor Fault Diagnosis Using Vibration and Acoustic Data

Sertac Kilickaya, Levent Eren

Purpose: The primary aim of this study is to enhance fault diagnosis in induction machines by leveraging the Pad\'e Approximant Neuron (PAON) model. While accelerometers and microphones are standard in motor condition monitoring, deep learning models with nonlinear neuron architectures offer promising improvements in diagnostic performance. This research investigates whether Pad\'e Approximant Neural Networks (Pad\'eNets) can outperform conventional Convolutional Neural Networks (CNNs) and Self-Organized Operational Neural Networks (Self-ONNs) in the diagnosis of electrical and mechanical faults from vibration and acoustic data. Methods: We evaluate and compare the diagnostic capabilities of three deep learning architectures: one-dimensional CNNs, Self-ONNs, and Pad\'eNets. These models are tested on the University of Ottawa's publicly available constant-speed induction motor datasets, which include both vibration and acoustic sensor data. The Pad\'eNet model is designed to introduce enhanced nonlinearity and is compatible with unbounded activation functions such as LeakyReLU. Results and Conclusion: Pad\'eNets consistently outperformed the baseline models, achieving diagnostic accuracies of 99.96%, 98.26%, 97.61%, and 98.33% for accelerometers 1, 2, 3, and the acoustic sensor, respectively. The enhanced nonlinearity of Pad\'eNets, together with their compatibility with unbounded activation functions, significantly improves fault diagnosis performance in induction motor condition monitoring.

comment: This version is the author's accepted manuscript. It has been peer-reviewed and accepted for publication in Journal of Vibration Engineering & Technologies. The final published version is available at https://doi.org/10.1007/s42417-025-02129-5

♻ ☆ TISDiSS: A Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation ICASSP 2026

Yongsheng Feng, Yuetonghui Xu, Jiehui Luo, Hongjia Liu, Xiaobing Li, Feng Yu, Wei Li

Source separation is a fundamental task in speech, music, and audio processing, and it also provides cleaner and larger data for training generative models. However, improving separation performance in practice often depends on increasingly large networks, inflating training and deployment costs. Motivated by recent advances in inference-time scaling for generative modeling, we propose Training-Time and Inference-Time Scalable Discriminative Source Separation (TISDiSS), a unified framework that integrates early-split multi-loss supervision, shared-parameter design, and dynamic inference repetitions. TISDiSS enables flexible speed-performance trade-offs by adjusting inference depth without retraining additional models. We further provide systematic analyses of architectural and training choices and show that training with more inference repetitions improves shallow-inference performance, benefiting low-latency applications. Experiments on standard speech separation benchmarks demonstrate state-of-the-art performance with a reduced parameter count, establishing TISDiSS as a scalable and practical framework for adaptive source separation. Code is available at https://github.com/WingSingFung/TISDiSS.

comment: Submitted to ICASSP 2026.(C) 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work

♻ ☆ AsynFusion: Towards Asynchronous Latent Consistency Models for Decoupled Whole-Body Audio-Driven Avatars

Tianbao Zhang, Jian Zhao, Yuer Li, Zheng Zhu, Ping Hu, Zhaoxin Fan, Wenjun Wu, Xuelong Li

Whole-body audio-driven avatar pose and expression generation is a critical task for creating lifelike digital humans and enhancing the capabilities of interactive virtual agents, with wide-ranging applications in virtual reality, digital entertainment, and remote communication. Existing approaches often generate audio-driven facial expressions and gestures independently, which introduces a significant limitation: the lack of seamless coordination between facial and gestural elements, resulting in less natural and cohesive animations. To address this limitation, we propose AsynFusion, a novel framework that leverages diffusion transformers to achieve harmonious expression and gesture synthesis. The proposed method is built upon a dual-branch DiT architecture, which enables the parallel generation of facial expressions and gestures. Within the model, we introduce a Cooperative Synchronization Module to facilitate bidirectional feature interaction between the two modalities, and an Asynchronous LCM Sampling strategy to reduce computational overhead while maintaining high-quality outputs. Extensive experiments demonstrate that AsynFusion achieves state-of-the-art performance in generating real-time, synchronized whole-body animations, consistently outperforming existing methods in both quantitative and qualitative evaluations.

comment: 15pages, conference

♻ ☆ ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

Mohammad Javad Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery

Existing Persian speech datasets are typically smaller than their English counterparts, which creates a key limitation for developing Persian speech technologies. We address this gap by introducing ParsVoice, the largest Persian speech corpus designed specifically for text-to-speech(TTS) applications. We created an automated pipeline that transforms raw audiobook content into TTS-ready data, incorporating components such as a BERT-based sentence completion detector, a binary search boundary optimization method for precise audio-text alignment, and audio-text quality assessment frameworks tailored to Persian. The pipeline processes 2,000 audiobooks, yielding 3,526 hours of clean speech, which was further filtered into a 1,804-hour high-quality subset suitable for TTS, featuring more than 470 speakers. To validate the dataset, we fine-tuned XTTS for Persian, achieving a naturalness Mean Opinion Score (MOS) of 3.6/5 and a Speaker Similarity Mean Opinion Score (SMOS) of 4.0/5 demonstrating ParsVoice's effectiveness for training multi-speaker TTS systems. ParsVoice is the largest high-quality Persian speech dataset, offering speaker diversity and audio quality comparable to major English corpora. The complete dataset has been made publicly available to accelerate the development of Persian speech technologies. The ParsVoice dataset is publicly available at: https://huggingface.co/datasets/MohammadJRanjbar/ParsVoice.

♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025

Huang-Cheng Chou, Haibin Wu, Hung-yi Lee, Chi-Chun Lee

Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.

comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025

♻ ☆ MRSAudio: A Large-Scale Multimodal Recorded Spatial Audio Dataset with Refined Annotations

Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Xintong Hu, Yu Zhang, Li Tang, Rui Yang, Han Wang, Zongbao Zhang, Yuhan Wang, Yixuan Chen, Hankun Xu, Ke Xu, Pengfei Fan, Zhetao Chen, Yanhao Yu, Qiange Huang, Fei Wu, Zhou Zhao

Humans rely on multisensory integration to perceive spatial environments, where auditory cues enable sound source localization in three-dimensional space. Despite the critical role of spatial audio in immersive technologies such as VR/AR, most existing multimodal datasets provide only monaural audio, which limits the development of spatial audio generation and understanding. To address these challenges, we introduce MRSAudio, a large-scale multimodal spatial audio dataset designed to advance research in spatial audio understanding and generation. MRSAudio spans four distinct components: MRSLife, MRSSpeech, MRSMusic, and MRSSing, covering diverse real-world scenarios. The dataset includes synchronized binaural and ambisonic audio, exocentric and egocentric video, motion trajectories, and fine-grained annotations such as transcripts, phoneme boundaries, lyrics, scores, and prompts. To demonstrate the utility and versatility of MRSAudio, we establish five foundational tasks: audio spatialization, and spatial text to speech, spatial singing voice synthesis, spatial music generation and sound event localization and detection. Results show that MRSAudio enables high-quality spatial modeling and supports a broad range of spatial audio research. Demos and dataset access are available at https://mrsaudio.github.io.

comment: 24 pages

♻ ☆ A Fast and Lightweight Model for Causal Audio-Visual Speech Separation ECAI 2025

Wendi Sang, Kai Li, Runxuan Yang, Jianqiang Huang, Xiaolin Hu

Audio-visual speech separation (AVSS) aims to extract a target speech signal from a mixed signal by leveraging both auditory and visual (lip movement) cues. However, most existing AVSS methods exhibit complex architectures and rely on future context, operating offline, which renders them unsuitable for real-time applications. Inspired by the pipeline of RTFSNet, we propose a novel streaming AVSS model, named Swift-Net, which enhances the causal processing capabilities required for real-time applications. Swift-Net adopts a lightweight visual feature extraction module and an efficient fusion module for audio-visual integration. Additionally, Swift-Net employs Grouped SRUs to integrate historical information across different feature spaces, thereby improving the utilization efficiency of historical information. We further propose a causal transformation template to facilitate the conversion of non-causal AVSS models into causal counterparts. Experiments on three standard benchmark datasets (LRS2, LRS3, and VoxCeleb2) demonstrated that under causal conditions, our proposed Swift-Net exhibited outstanding performance, highlighting the potential of this method for processing speech in complex environments.

comment: Accepted by ECAI 2025

♻ ☆ PAL: Probing Audio Encoders via LLMs - Audio Information Transfer into LLMs

Tony Alex, Wish Suharitdamrong, Sara Atito, Armin Mustafa, Philip J. B. Jackson, Imran Razzak, Muhammad Awais

Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects the audio encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former), then prepends or inserts them to the text tokens. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL introduces audio representations solely via the attention mechanism within different layers of the LLM, bypassing its feedforward module. LAL encodes rich audio semantics at an appropriate level of abstraction for integration into different blocks of LLMs. Our design significantly reduces computational overhead compared to existing integration approaches. Observing with Whisper that the speech encoder benefits from PLITS integration, we propose an audio encoder aware approach for efficiently Probing Audio encoders via LLM (PAL), which employs PLITS integration for Whisper and LAL for general audio encoders. Under an identical training curriculum, LAL consistently maintains performance or outperforms existing integration approaches across multiple base LLMs and tasks. For general audio tasks, LAL improvement is up to 30% over a strong PLITS baseline while reducing memory usage by up to 64.1% and increasing throughput by up to 247.5%. Furthermore, for general audio-music-speech LLM, PAL performs on par with a fully PLITS integration-based system but with substantially improved computational and memory efficiency. Project page: https://ta012.github.io/PAL/

comment: 17 pages, 3 figures

☆ I-DCCRN-VAE: An Improved Deep Representation Learning Framework for Complex VAE-based Single-channel Speech Enhancement

Jiatong Li, Simon Doclo

Recently, a complex variational autoencoder (VAE)-based single-channel speech enhancement system based on the DCCRN architecture has been proposed. In this system, a noise suppression VAE (NSVAE) learns to extract clean speech representations from noisy speech using pretrained clean speech and noise VAEs with skip connections. In this paper, we improve DCCRN-VAE by incorporating three key modifications: 1) removing the skip connections in the pretrained VAEs to encourage more informative speech and noise latent representations; 2) using $\beta$-VAE in pretraining to better balance reconstruction and latent space regularization; and 3) a NSVAE generating both speech and noise latent representations. Experiments show that the proposed system achieves comparable performance as the DCCRN and DCCRN-VAE baselines on the matched DNS3 dataset but outperforms the baselines on mismatched datasets (WSJ0-QUT, Voicebank-DEMEND), demonstrating improved generalization ability. In addition, an ablation study shows that a similar performance can be achieved with classical fine-tuning instead of adversarial training, resulting in a simpler training pipeline.

☆ A Phase Synthesizer for Decorrelation to Improve Acoustic Feedback Cancellation

Klaus Linhard, Philipp Bulling

Undesired acoustic feedback is a known issue in communication systems, such as speech in-car communication, public address systems, or hearing aids. Without additional precautions, there is a high risk that the adaptive filter - intended to cancel the feedback path - also suppresses parts of the desired signal. One solution is to decorrelate the loudspeaker and microphone signals. In this work, we combine the two decorrelation approaches frequency shifting and phase modulation in a unified framework: a so-called \textit{phase synthesizer}, implemented in a discrete Fourier transform (DFT) filter bank. Furthermore, we extend the phase modulation technique using variable delay lines, as known from vibrato and chorus effects. We demonstrate the benefits of the proposed phase synthesizer using an example from speech in-car communication, employing an adaptive frequency-domain Kalman filter. Improvements in system stability, speech quality measured by perceptual evaluation of speech quality (PESQ) are presented.

☆ DeePAQ: A Perceptual Audio Quality Metric Based On Foundational Models and Weakly Supervised Learning

Guanxin Jiang, Andreas Brendel, Pablo M. Delgado, Jürgen Herre

This paper presents the Deep learning-based Perceptual Audio Quality metric (DeePAQ) for evaluating general audio quality. Our approach leverages metric learning together with the music foundation model MERT, guided by surrogate labels, to construct an embedding space that captures distortion intensity in general audio. To the best of our knowledge, DeePAQ is the first in the general audio quality domain to leverage weakly supervised labels and metric learning for fine-tuning a music foundation model with Low-Rank Adaptation (LoRA), a direction not yet explored by other state-of-the-art methods. We benchmark the proposed model against state-of-the-art objective audio quality metrics across listening tests spanning audio coding and source separation. Results show that our method surpasses existing metrics in detecting coding artifacts and generalizes well to unseen distortions such as source separation, highlighting its robustness and versatility.

comment: 5 pages, 2 figures

☆ DiSTAR: Diffusion over a Scalable Token Autoregressive Representation for Speech Generation

Yakun Song, Xiaobin Zhuang, Jiawei Chen, Zhikang Niu, Guanrou Yang, Chenpeng Du, Zhuo Chen, Yuping Wang, Yuxuan Wang, Xie Chen

Recent attempts to interleave autoregressive (AR) sketchers with diffusion-based refiners over continuous speech representations have shown promise, but they remain brittle under distribution shift and offer limited levers for controllability. We introduce DISTAR, a zero-shot text-to-speech framework that operates entirely in a discrete residual vector quantization (RVQ) code space and tightly couples an AR language model with a masked diffusion model, without forced alignment or a duration predictor. Concretely, DISTAR drafts block-level RVQ tokens with an AR language model and then performs parallel masked-diffusion infilling conditioned on the draft to complete the next block, yielding long-form synthesis with blockwise parallelism while mitigating classic AR exposure bias. The discrete code space affords explicit control at inference: DISTAR produces high-quality audio under both greedy and sample-based decoding using classifier-free guidance, supports trade-offs between robustness and diversity, and enables variable bit-rate and controllable computation via RVQ layer pruning at test time. Extensive experiments and ablations demonstrate that DISTAR surpasses state-of-the-art zero-shot TTS systems in robustness, naturalness, and speaker/style consistency, while maintaining rich output diversity. Audio samples are provided on https://anonymous.4open.science/w/DiSTAR_demo.

☆ Audio Palette: A Diffusion Transformer with Multi-Signal Conditioning for Controllable Foley Synthesis

Junnuo Wang

Recent advances in diffusion-based generative models have enabled high-quality text-to-audio synthesis, but fine-grained acoustic control remains a significant challenge in open-source research. We present Audio Palette, a diffusion transformer (DiT) based model that extends the Stable Audio Open architecture to address this "control gap" in controllable audio generation. Unlike prior approaches that rely solely on semantic conditioning, Audio Palette introduces four time-varying control signals: loudness, pitch, spectral centroid, and timbre, for precise and interpretable manipulation of acoustic features. The model is efficiently adapted for the nuanced domain of Foley synthesis using Low-Rank Adaptation (LoRA) on a curated subset of AudioSet, requiring only 0.85 percent of the original parameters to be trained. Experiments demonstrate that Audio Palette achieves fine-grained, interpretable control of sound attributes. Crucially, it accomplishes this novel controllability while maintaining high audio quality and strong semantic alignment to text prompts, with performance on standard metrics such as Frechet Audio Distance (FAD) and LAION-CLAP scores remaining comparable to the original baseline model. We provide a scalable, modular pipeline for audio research, emphasizing sequence-based conditioning, memory efficiency, and a three-scale classifier-free guidance mechanism for nuanced inference-time control. This work establishes a robust foundation for controllable sound design and performative audio synthesis in open-source settings, enabling a more artist-centric workflow.

comment: Accepted for publication in the Journal of Artificial Intelligence Research (JAIR), Vol. 3 No. 2, December 2025

☆ FakeMark: Deepfake Speech Attribution With Watermarked Artifacts

Wanying Ge, Xin Wang, Junichi Yamagishi

Deepfake speech attribution remains challenging for existing solutions. Classifier-based solutions often fail to generalize to domain-shifted samples, and watermarking-based solutions are easily compromised by distortions like codec compression or malicious removal attacks. To address these issues, we propose FakeMark, a novel watermarking framework that injects artifact-correlated watermarks associated with deepfake systems rather than pre-assigned bitstring messages. This design allows a detector to attribute the source system by leveraging both injected watermark and intrinsic deepfake artifacts, remaining effective even if one of these cues is elusive or removed. Experimental results show that FakeMark improves generalization to cross-dataset samples where classifier-based solutions struggle and maintains high accuracy under various distortions where conventional watermarking-based solutions fail.

☆ Switchboard-Affect: Emotion Perception Labels from Conversational Speech

Amrit Romana, Jaya Narain, Tien Dung Tran, Andrea Davis, Jason Fong, Ramya Rasipuram, Vikramjit Mitra

Understanding the nuances of speech emotion dataset curation and labeling is essential for assessing speech emotion recognition (SER) model potential in real-world applications. Most training and evaluation datasets contain acted or pseudo-acted speech (e.g., podcast speech) in which emotion expressions may be exaggerated or otherwise intentionally modified. Furthermore, datasets labeled based on crowd perception often lack transparency regarding the guidelines given to annotators. These factors make it difficult to understand model performance and pinpoint necessary areas for improvement. To address this gap, we identified the Switchboard corpus as a promising source of naturalistic conversational speech, and we trained a crowd to label the dataset for categorical emotions (anger, contempt, disgust, fear, sadness, surprise, happiness, tenderness, calmness, and neutral) and dimensional attributes (activation, valence, and dominance). We refer to this label set as Switchboard-Affect (SWB-Affect). In this work, we present our approach in detail, including the definitions provided to annotators and an analysis of the lexical and paralinguistic cues that may have played a role in their perception. In addition, we evaluate state-of-the-art SER models, and we find variable performance across the emotion categories with especially poor generalization for anger. These findings underscore the importance of evaluation with datasets that capture natural affective variations in speech. We release the labels for SWB-Affect to enable further analysis in this domain.

comment: 2025 13th International Conference on Affective Computing and Intelligent Interaction (ACII) https://github.com/apple/ml-switchboard-affect

☆ Continuous-Token Diffusion for Speaker-Referenced TTS in Multimodal LLMs

Xinlu He, Swayambhu Nath Ray, Harish Mallidi, Jia-Hong Huang, Ashwin Bellur, Chander Chandak, M. Maruf, Venkatesh Ravichandran

Unified architectures in multimodal large language models (MLLM) have shown promise in handling diverse tasks within a single framework. In the text-to-speech (TTS) task, current MLLM-based approaches rely on discrete token representations, which disregard the inherently continuous nature of speech and can lead to loss of fine-grained acoustic information.In this work, we investigate the TTS within the MLLM paradigm using continuous speech representations. We design a dual-head architecture and implement two complementary training strategies for a robust model. (1) A diffusion head generating continuous speech representations is added on the MLLM, which is on frame-level and strictly autoregressive. (2) The original language model head is retained to preserve multitask capability and to control the start and end of speech synthesis. (3) Masked training is employed to address exposure bias in autoregressive decoding. (4) To stabilize optimization, we propose a two-stage scheme where the LM is frozen in the second stage, ensuring the diffusion head learns from a fixed input distribution. Evaluations on LibriSpeech(PC) test-clean show that our approach achieves state-of-the-art autoregressive performance, with a WER of 1.95%, speaker similarity of 0.54, and UTMOS of 4.00. The two-stage training yields a 46% relative WER reduction over the one-stage training baseline. These results highlight the effectiveness of combining autoregressive modeling with continuous-token diffusion, supported by a two-stage training procedure.

☆ HyWA: Hypernetwork Weight Adapting Personalized Voice Activity Detection

Mahsa Ghazvini Nejad, Hamed Jafarzadeh Asl, Amin Edraki, Mohammadreza Sadeghi, Masoud Asgharian, Yuanhao Yu, Vahid Partovi Nia

Personalized Voice Activity Detection (PVAD) systems activate only in response to a specific target speaker by incorporating speaker embeddings from enrollment utterances. Unlike existing methods that require architectural changes, such as FiLM layers, our approach employs a hypernetwork to modify the weights of a few selected layers within a standard voice activity detection (VAD) model. This enables speaker conditioning without changing the VAD architecture, allowing the same VAD model to adapt to different speakers by updating only a small subset of the layers. We propose HyWA-PVAD, a hypernetwork weight adaptation method, and evaluate it against multiple baseline conditioning techniques. Our comparison shows consistent improvements in PVAD performance. HyWA also offers practical advantages for deployment by preserving the core VAD architecture. Our new approach improves the current conditioning techniques in two ways: i) increases the mean average precision, ii) simplifies deployment by reusing the same VAD architecture.

comment: Mahsa Ghazvini Nejad and Hamed Jafarzadeh Asl contributed equally to this work

☆ Adaptive vector steering: A training-free, layer-wise intervention for hallucination mitigation in large audio and multimodal models ICASSP 2026

Tsung-En Lin, Kuan-Yi Lee, Hung-Yi Lee

Large Audio-Language Models and Multi-Modal Large Language Models have demonstrated strong capabilities in tasks such as Audio Question Answering (AQA), Audio Captioning, and Automatic Speech Recognition (ASR). However, there is growing evidence that these models can hallucinate about the content of the audio. To address this issue, we probe the models' internal states and propose Adaptive Vector Steering (AVS), a method that better grounds generation in audio content. We also identify a strong correlation between output correctness and internal representations. Experiments show consistent performance gains across two models and two benchmarks. On the Audio Hallucination QA dataset, our method boosts the F1-score of Gemma from 0.550 to 0.619 and Qwen from 0.626 to 0.632. Furthermore, our method increases the accuracy of Qwen on MMAU from 0.548 to 0.592, marking an 8% relative increase. To the best of our knowledge, this is the first work to apply vector steering to mitigate hallucination in audio.

comment: Note: This preprint is a version of the paper submitted to ICASSP 2026. The author list here includes contributors who provided additional supervision and guidance. The official ICASSP submission may differ slightly in author composition

♻ ☆ Investigating Faithfulness in Large Audio Language Models

Lovenya Jain, Pooneh Mousavi, Mirco Ravanelli, Cem Subakan

Faithfulness measures whether chain-of-thought (CoT) representations accurately reflect a model's decision process and can be used as reliable explanations. Prior work has shown that CoTs from text-based LLMs are often unfaithful. This question has not been explored for large audio-language models (LALMs), where faithfulness is critical for safety-sensitive applications. Reasoning in LALMs is also more challenging, as models must first extract relevant clues from audio before reasoning over them. In this paper, we investigate the faithfulness of CoTs produced by several LALMs by applying targeted interventions, including paraphrasing, filler token injection, early answering, and introducing mistakes, on two challenging reasoning datasets: SAKURA and MMAR. After going through the aforementioned interventions across several datasets and tasks, our experiments suggest that, LALMs generally produce CoTs that appear to be faithful to their underlying decision processes.

♻ ☆ TISDiSS: A Training-Time and Inference-Time Scalable Framework for Discriminative Source Separation ICASSP 2026

Yongsheng Feng, Yuetonghui Xu, Jiehui Luo, Hongjia Liu, Xiaobing Li, Feng Yu, Wei Li

Source separation is a fundamental task in speech, music, and audio processing, and it also provides cleaner and larger data for training generative models. However, improving separation performance in practice often depends on increasingly large networks, inflating training and deployment costs. Motivated by recent advances in inference-time scaling for generative modeling, we propose Training-Time and Inference-Time Scalable Discriminative Source Separation (TISDiSS), a unified framework that integrates early-split multi-loss supervision, shared-parameter design, and dynamic inference repetitions. TISDiSS enables flexible speed-performance trade-offs by adjusting inference depth without retraining additional models. We further provide systematic analyses of architectural and training choices and show that training with more inference repetitions improves shallow-inference performance, benefiting low-latency applications. Experiments on standard speech separation benchmarks demonstrate state-of-the-art performance with a reduced parameter count, establishing TISDiSS as a scalable and practical framework for adaptive source separation. Code is available at https://github.com/WingSingFung/TISDiSS.

comment: Submitted to ICASSP 2026.(C) 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work

♻ ☆ AsynFusion: Towards Asynchronous Latent Consistency Models for Decoupled Whole-Body Audio-Driven Avatars

Tianbao Zhang, Jian Zhao, Yuer Li, Zheng Zhu, Ping Hu, Zhaoxin Fan, Wenjun Wu, Xuelong Li

Whole-body audio-driven avatar pose and expression generation is a critical task for creating lifelike digital humans and enhancing the capabilities of interactive virtual agents, with wide-ranging applications in virtual reality, digital entertainment, and remote communication. Existing approaches often generate audio-driven facial expressions and gestures independently, which introduces a significant limitation: the lack of seamless coordination between facial and gestural elements, resulting in less natural and cohesive animations. To address this limitation, we propose AsynFusion, a novel framework that leverages diffusion transformers to achieve harmonious expression and gesture synthesis. The proposed method is built upon a dual-branch DiT architecture, which enables the parallel generation of facial expressions and gestures. Within the model, we introduce a Cooperative Synchronization Module to facilitate bidirectional feature interaction between the two modalities, and an Asynchronous LCM Sampling strategy to reduce computational overhead while maintaining high-quality outputs. Extensive experiments demonstrate that AsynFusion achieves state-of-the-art performance in generating real-time, synchronized whole-body animations, consistently outperforming existing methods in both quantitative and qualitative evaluations.

comment: 15pages, conference

♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025

Huang-Cheng Chou, Haibin Wu, Hung-yi Lee, Chi-Chun Lee

Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.

comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025

♻ ☆ A Fast and Lightweight Model for Causal Audio-Visual Speech Separation ECAI 2025

Wendi Sang, Kai Li, Runxuan Yang, Jianqiang Huang, Xiaolin Hu

Audio-visual speech separation (AVSS) aims to extract a target speech signal from a mixed signal by leveraging both auditory and visual (lip movement) cues. However, most existing AVSS methods exhibit complex architectures and rely on future context, operating offline, which renders them unsuitable for real-time applications. Inspired by the pipeline of RTFSNet, we propose a novel streaming AVSS model, named Swift-Net, which enhances the causal processing capabilities required for real-time applications. Swift-Net adopts a lightweight visual feature extraction module and an efficient fusion module for audio-visual integration. Additionally, Swift-Net employs Grouped SRUs to integrate historical information across different feature spaces, thereby improving the utilization efficiency of historical information. We further propose a causal transformation template to facilitate the conversion of non-causal AVSS models into causal counterparts. Experiments on three standard benchmark datasets (LRS2, LRS3, and VoxCeleb2) demonstrated that under causal conditions, our proposed Swift-Net exhibited outstanding performance, highlighting the potential of this method for processing speech in complex environments.

comment: Accepted by ECAI 2025

♻ ☆ PAL: Probing Audio Encoders via LLMs - Audio Information Transfer into LLMs

Tony Alex, Wish Suharitdamrong, Sara Atito, Armin Mustafa, Philip J. B. Jackson, Imran Razzak, Muhammad Awais

Integration of audio perception into large language models (LLMs) is an emerging research area for enabling machine listening applications, yet efficient transfer of rich audio semantics from audio encoders to LLMs remains underexplored. The most widely used integration paradigm projects the audio encoder output tokens into the LLM input space (e.g., via an MLP or a Q-Former), then prepends or inserts them to the text tokens. We refer to this generic scheme as Prepend to the LLM's input token space (PLITS) integration. We propose an efficient alternative, Lightweight Audio LLM Integration (LAL). LAL introduces audio representations solely via the attention mechanism within different layers of the LLM, bypassing its feedforward module. LAL encodes rich audio semantics at an appropriate level of abstraction for integration into different blocks of LLMs. Our design significantly reduces computational overhead compared to existing integration approaches. Observing with Whisper that the speech encoder benefits from PLITS integration, we propose an audio encoder aware approach for efficiently Probing Audio encoders via LLM (PAL), which employs PLITS integration for Whisper and LAL for general audio encoders. Under an identical training curriculum, LAL consistently maintains performance or outperforms existing integration approaches across multiple base LLMs and tasks. For general audio tasks, LAL improvement is up to 30% over a strong PLITS baseline while reducing memory usage by up to 64.1% and increasing throughput by up to 247.5%. Furthermore, for general audio-music-speech LLM, PAL performs on par with a fully PLITS integration-based system but with substantially improved computational and memory efficiency. Project page: https://ta012.github.io/PAL/

comment: 17 pages, 3 figures

☆ BridgeCode: A Dual Speech Representation Paradigm for Autoregressive Zero-Shot Text-to-Speech Synthesis

Jingyuan Xing, Mingru Yang, Zhipeng Li, Xiaofen Xing, Xiangmin Xu

Autoregressive (AR) frameworks have recently achieved remarkable progress in zero-shot text-to-speech (TTS) by leveraging discrete speech tokens and large language model techniques. Despite their success, existing AR-based zero-shot TTS systems face two critical limitations: (i) an inherent speed-quality trade-off, as sequential token generation either reduces frame rates at the cost of expressiveness or enriches tokens at the cost of efficiency, and (ii) a text-oriented supervision mismatch, as cross-entropy loss penalizes token errors uniformly without considering the fine-grained acoustic similarity among adjacent tokens. To address these challenges, we propose BridgeTTS, a novel AR-TTS framework built upon the dual speech representation paradigm BridgeCode. BridgeTTS reduces AR iterations by predicting sparse tokens while reconstructing rich continuous features for high-quality synthesis. Joint optimization of token-level and feature-level objectives further enhances naturalness and intelligibility. Experiments demonstrate that BridgeTTS achieves competitive quality and speaker similarity while significantly accelerating synthesis. Speech demos are available at https://test1562.github.io/demo/.

☆ Automatic Music Sample Identification with Multi-Track Contrastive Learning

Alain Riou, Joan Serrà, Yuki Mitsufuji

Sampling, the technique of reusing pieces of existing audio tracks to create new music content, is a very common practice in modern music production. In this paper, we tackle the challenging task of automatic sample identification, that is, detecting such sampled content and retrieving the material from which it originates. To do so, we adopt a self-supervised learning approach that leverages a multi-track dataset to create positive pairs of artificial mixes, and design a novel contrastive learning objective. We show that such method significantly outperforms previous state-of-the-art baselines, that is robust to various genres, and that scales well when increasing the number of noise songs in the reference database. In addition, we extensively analyze the contribution of the different components of our training pipeline and highlight, in particular, the need for high-quality separated stems for this task.

☆ Audio-Maestro: Enhancing Large Audio-Language Models with Tool-Augmented Reasoning

Kuan-Yi Lee, Tsung-En Lin, Hung-Yi Lee

Recent advancements in large multimodal models (LMMs) have shown strong capabilities in audio understanding. However, most systems rely solely on end-to-end reasoning, limiting interpretability and accuracy for tasks that require structured knowledge or specialized signal analysis. In this work, we present Audio-Maestro -- a tool-augmented audio reasoning framework that enables audio-language models to autonomously call external tools and integrate their timestamped outputs into the reasoning process. This design allows the model to analyze, transform, and interpret audio signals through specialized tools rather than relying solely on end-to-end inference. Experiments show that Audio-Maestro consistently improves general audio reasoning performance: Gemini-2.5-flash's average accuracy on MMAU-Test rises from 67.4% to 72.1%, DeSTA-2.5 from 58.3% to 62.8%, and GPT-4o from 60.8% to 63.9%. To our knowledge, Audio-Maestro is the first framework to integrate structured tool output into the large audio language model reasoning process.

comment: 9pages

☆ Diffusion-Link: Diffusion Probabilistic Model for Bridging the Audio-Text Modality Gap ICASSP 2026

KiHyun Nam, Jongmin Choi, Hyeongkeun Lee, Jungwoo Heo, Joon Son Chung

Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link

comment: 5 pages. Submitted to IEEE ICASSP 2026

☆ Perturbation Self-Supervised Representations for Cross-Lingual Emotion TTS: Stage-Wise Modeling of Emotion and Speaker

Cheng Gong, Chunyu Qiang, Tianrui Wang, Yu Jiang, Yuheng Lu, Ruihao Jing, Xiaoxiao Miao, Xiaolei Zhang, Longbiao Wang, Jianwu Dang

Cross-lingual emotional text-to-speech (TTS) aims to produce speech in one language that captures the emotion of a speaker from another language while maintaining the target voice's timbre. This process of cross-lingual emotional speech synthesis presents a complex challenge, necessitating flexible control over emotion, timbre, and language. However, emotion and timbre are highly entangled in speech signals, making fine-grained control challenging. To address this issue, we propose EMM-TTS, a novel two-stage cross-lingual emotional speech synthesis framework based on perturbed self-supervised learning (SSL) representations. In the first stage, the model explicitly and implicitly encodes prosodic cues to capture emotional expressiveness, while the second stage restores the timbre from perturbed SSL representations. We further investigate the effect of different speaker perturbation strategies-formant shifting and speaker anonymization-on the disentanglement of emotion and timbre. To strengthen speaker preservation and expressive control, we introduce Speaker Consistency Loss (SCL) and Speaker-Emotion Adaptive Layer Normalization (SEALN) modules. Additionally, we find that incorporating explicit acoustic features (e.g., F0, energy, and duration) alongside pretrained latent features improves voice cloning performance. Comprehensive multi-metric evaluations, including both subjective and objective measures, demonstrate that EMM-TTS achieves superior naturalness, emotion transferability, and timbre consistency across languages.

comment: Submitted to Expert Systems with Applications,11 pages

☆ VCB Bench: An Evaluation Benchmark for Audio-Grounded Large Language Model Conversational Agents

Jiliang Hu, Wenfu Wang, Zuchao Li, Chenxing Li, Yiyang Zhao, Hanzhao Li, Liqiang Zhang, Meng Yu, Dong Yu

Recent advances in large audio language models (LALMs) have greatly enhanced multimodal conversational systems. However, existing benchmarks remain limited -- they are mainly English-centric, rely on synthetic speech, and lack comprehensive, discriminative evaluation across multiple dimensions. To address these gaps, we present Voice Chat Bot Bench (VCB Bench) -- a high-quality Chinese benchmark built entirely on real human speech. VCB Bench evaluates LALMs from three complementary perspectives: instruction following (including speech-level control beyond text commands), knowledge understanding (general knowledge, reasoning, and daily dialogue), and robustness (stability under perturbations in content, environment, and speaker traits). Experiments on representative LALMs reveal notable performance gaps and highlight future directions for improvement. VCB Bench provides a reproducible and fine-grained evaluation framework, offering standardized methodology and practical insights for advancing Chinese voice conversational models.

comment: 20 pages, 5 figures

☆ MSRBench: A Benchmarking Dataset for Music Source Restoration

Yongyi Zang, Jiarui Hai, Wanying Ge, Qiuqiang Kong, Zheqi Dai, Helin Wang, Yuki Mitsufuji, Mark D. Plumbley

Music Source Restoration (MSR) extends source separation to realistic settings where signals undergo production effects (equalization, compression, reverb) and real-world degradations, with the goal of recovering the original unprocessed sources. Existing benchmarks cannot measure restoration fidelity: synthetic datasets use unprocessed stems but unrealistic mixtures, while real production datasets provide only already-processed stems without clean references. We present MSRBench, the first benchmark explicitly designed for MSR evaluation. MSRBench contains raw stem-mixture pairs across eight instrument classes, where mixtures are produced by professional mixing engineers. These raw-processed pairs enable direct evaluation of both separation accuracy and restoration fidelity. Beyond controlled studio conditions, the mixtures are augmented with twelve real-world degradations spanning analog artifacts, acoustic environments, and lossy codecs. Baseline experiments with U-Net and BSRNN achieve SI-SNR of -37.8 dB and -23.4 dB respectively, with perceptual quality (FAD CLAP) around 0.7-0.8, demonstrating substantial room for improvement and the need for restoration-specific architectures.

☆ Unify Variables in Neural Scaling Laws for General Audio Representations via Embedding Effective Rank

Xuyao Deng, Yanjie Sun, Yong Dou, Kele Xu

Scaling laws have profoundly shaped our understanding of model performance in computer vision and natural language processing, yet their application to general audio representation learning remains underexplored. A key challenge lies in the multifactorial nature of general audio representation-representation quality is jointly influenced by variables such as audio length, embedding dimensionality, model depth, model architecture, data volume, etc., many of which are difficult to isolate or express analytically. In this work, we present a systematic study of scaling laws for general audio representations by utilizing embedding effective rank (RankMe) as a unifying metric that encapsulates the impact of diverse variables on representation quality. RankMe enables a label-free, information-theoretic quantification of audio embeddings, allowing us to examine scaling behaviors across a wide hyper-parameter space, including model size, training data volume, computational budget, architectural configurations, etc. Our empirical findings reveal a consistent power-law relationship between RankMe and representation quality, suggesting that embedding effective rank serves as a reliable proxy for assessing and predicting model performance in audio representation learning. This work not only validates the applicability of classical scaling principles to the general audio domain but also offers a theoretically grounded and empirically robust framework for guiding future model scaling strategies in audio foundation models.

☆ UALM: Unified Audio Language Model for Understanding, Generation and Reasoning

Jinchuan Tian, Sang-gil Lee, Zhifeng Kong, Sreyan Ghosh, Arushi Goel, Chao-Han Huck Yang, Wenliang Dai, Zihan Liu, Hanrong Ye, Shinji Watanabe, Mohammad Shoeybi, Bryan Catanzaro, Rafael Valle, Wei Ping

Recent advances in the audio language modeling (ALM) domain tackle audio understanding and text-to-audio generation as separate tasks. Very few studies attempt to unify these tasks -- an essential step toward advanced multimodal reasoning. This paper introduces U}nified Audio Language Model (UALM), which aims to unify audio understanding, text-to-audio generation, and multimodal reasoning in a single model. To achieve this goal, we first present UALM-Gen, a text-to-audio language model that directly predicts audio tokens and is comparable to state-of-the-art diffusion-based models. We then demonstrate, using proper data blending, training recipes, and inference techniques, that our single UALM model matches the quality of state-of-the-art specialized models in audio understanding, text-to-audio generation, and text reasoning. Furthermore, we present UALM-Reason, a multimodal reasoning model that utilizes both text and audio in the intermediate thinking steps to facilitate complex generation tasks. To our knowledge, this is the first demonstration in audio research of cross-modal generative reasoning, with its effectiveness confirmed by subjective evaluations.

☆ Audio-Guided Visual Perception for Audio-Visual Navigation

Yi Wang, Yinfeng Yu, Fuchun Sun, Liejun Wang, Wendong Zheng

Audio-Visual Embodied Navigation aims to enable agents to autonomously navigate to sound sources in unknown 3D environments using auditory cues. While current AVN methods excel on in-distribution sound sources, they exhibit poor cross-source generalization: navigation success rates plummet and search paths become excessively long when agents encounter unheard sounds or unseen environments. This limitation stems from the lack of explicit alignment mechanisms between auditory signals and corresponding visual regions. Policies tend to memorize spurious \enquote{acoustic fingerprint-scenario} correlations during training, leading to blind exploration when exposed to novel sound sources. To address this, we propose the AGVP framework, which transforms sound from policy-memorable acoustic fingerprint cues into spatial guidance. The framework first extracts global auditory context via audio self-attention, then uses this context as queries to guide visual feature attention, highlighting sound-source-related regions at the feature level. Subsequent temporal modeling and policy optimization are then performed. This design, centered on interpretable cross-modal alignment and region reweighting, reduces dependency on specific acoustic fingerprints. Experimental results demonstrate that AGVP improves both navigation efficiency and robustness while achieving superior cross-scenario generalization on previously unheard sounds.

comment: Main paper (6 pages). Accepted for publication by International Conference on Virtual Reality and Visualization 2025 (ICVRV 2025)

☆ Gelina: Unified Speech and Gesture Synthesis via Interleaved Token Prediction

Téo Guichoux, Théodor Lemerle, Shivam Mehta, Jonas Beskow, Gustave Eje Henter, Laure Soulier, Catherine Pelachaud, Nicolas Obin

Human communication is multimodal, with speech and gestures tightly coupled, yet most computational methods for generating speech and gestures synthesize them sequentially, weakening synchrony and prosody alignment. We introduce Gelina, a unified framework that jointly synthesizes speech and co-speech gestures from text using interleaved token sequences in a discrete autoregressive backbone, with modality-specific decoders. Gelina supports multi-speaker and multi-style cloning and enables gesture-only synthesis from speech inputs. Subjective and objective evaluations demonstrate competitive speech quality and improved gesture generation over unimodal baselines.

comment: 5 pages

♻ ☆ MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction NeurIPS 2025

Yunkee Chae, Kyogu Lee

We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.

comment: Accepted by NeurIPS 2025

♻ ☆ GRAM: Spatial general-purpose audio representation models for real-world applications

Goksenin Yuksel, Marcel van Gerven, Kiki van der Heijden

Although audio foundations models have seen great progress on a wide variety of tasks, their application in real-world acoustic environments with reverberation and noise has been less successful. Moreover, as audio foundation models are typically trained on dry, single-channel audio clips, the inherent spatial nature of real-world sound scenes is overlooked and tasks involving sound localization ruled out. To address these limitations, we propose GRAM: a General-purpose Real-world Audio Model utilizing a multi-channel masked auto-encoder approach to efficiently learn spatial audio representations from high-quality simulated real-world scenes. To evaluate the performance of GRAM and other audio foundation models in real-world sound scenes, we release Nat-HEAR: A naturalistic version of the HEAR benchmark suite comprising a simulated real-world version, as well as two new sound localization tasks. We show that the performance of GRAM surpasses all state-of-the-art self-supervised audio foundation models and speech models on both HEAR and Nat-HEAR, while using only a fraction of the training data. GRAM also showcases state-of-the-art localization performance, surpassing even supervised sound localization approaches, and can be flexibly applied either to a two-channel, binaural sound format or a four-channel, Ambisonics format. Validating GRAM's performance on real-world sound recordings demonstrates robust transfer to real-world scenes. Taken together, GRAM presents a significant advancement towards robust, spatial audio foundation models for real-world applications.

comment: Still under review

♻ ☆ WavJEPA: Semantic learning unlocks robust audio foundation models for raw waveforms

Goksenin Yuksel, Pierre Guetschel, Michael Tangermann, Marcel van Gerven, Kiki van der Heijden

Learning audio representations from raw waveforms overcomes key limitations of spectrogram-based audio representation learning, such as the long latency of spectrogram computation and the loss of phase information. Yet, while self-supervised speech representation learning from raw waveforms has been remarkably successful, these approaches have not achieved similar feats for general-purpose audio representation learning from waveforms. Here, we propose WavJEPA, a waveform-based version of the Joint-Embedding Predictive Architecture. WavJEPA leverages high-level semantic representation learning to tackle the shortcomings of representation learning at the speech unit or token level. We show that this approach substantially outperforms state-of-the-art time-domain audio foundation models across a wide variety of downstream benchmark tasks, while requiring considerably fewer computational resources. Additionally, to overcome the performance drop that time-domain models typically exhibit in noisy and reverberant real-world acoustic environments, we present WavJEPA-Nat. WavJEPA-Nat is a multi-channel extension of the WavJEPA architecture trained on simulated naturalistic scenes. We find that WavJEPA-Nat is highly robust to reverberation and noise. These results highlight the feasibility and computational efficiency of general-purpose audio representation learning from raw waveforms, showcasing the potential for low-latency, robust time-domain audio foundation models for real-world applications.

comment: Still under review

♻ ☆ Detecting and Mitigating Insertion Hallucination in Video-to-Audio Generation

Liyang Chen, Hongkai Chen, Yujun Cai, Sifan Li, Qingwen Ye, Yiwei Wang

Video-to-Audio generation has made remarkable strides in automatically synthesizing sound for video. However, existing evaluation metrics, which focus on semantic and temporal alignment, overlook a critical failure mode: models often generate acoustic events, particularly speech and music, that have no corresponding visual source. We term this phenomenon Insertion Hallucination and identify it as a systemic risk driven by dataset biases, such as the prevalence of off-screen sounds, that remains completely undetected by current metrics. To address this challenge, we first develop a systematic evaluation framework that employs a majority-voting ensemble of multiple audio event detectors. We also introduce two novel metrics to quantify the prevalence and severity of this issue: IH@vid (the fraction of videos with hallucinations) and IH@dur (the fraction of hallucinated duration). Building on this, we propose Posterior Feature Correction, a novel training-free inference-time method that mitigates IH. PFC operates in a two-pass process: it first generates an initial audio output to detect hallucinated segments, and then regenerates the audio after masking the corresponding video features at those timestamps. Experiments on several mainstream V2A benchmarks first reveal that state-of-the-art models suffer from severe IH. In contrast, our PFC method reduces both the prevalence and duration of hallucinations by over 50\% on average, without degrading, and in some cases even improving, conventional metrics for audio quality and temporal synchronization. Our work is the first to formally define, systematically measure, and effectively mitigate Insertion Hallucination, paving the way for more reliable and faithful V2A models.

♻ ☆ Joint Source-Environment Adaptation of Data-Driven Underwater Acoustic Source Ranging Based on Model Uncertainty

Dariush Kari, Hari Vishnu, Andrew C. Singer

Adapting pre-trained deep learning models to new and unknown environments remains a major challenge in underwater acoustic localization. We show that although the performance of pre-trained models suffers from mismatch between the training and test data, they generally exhibit a higher uncertainty in environments where there is more mismatch. Additionally, in the presence of environmental mismatch, spurious peaks can appear in the output of classification-based localization approaches, which inspires us to define and use a method to quantify the "implied uncertainty" based on the number of model output peaks. Leveraging this notion of implied uncertainty, we partition the test samples into sets with more certain and less certain samples, and implement a method to adapt the model to new environments by using the certain samples to improve the labeling for uncertain samples, which helps to adapt the model. Thus, using this efficient method for model uncertainty quantification, we showcase an innovative approach to adapt a pre-trained model to unseen underwater environments at test time. This eliminates the need for labeled data from the target environment or the original training data. This adaptation is enhanced by integrating an independent estimate based on the received signal energy. We validate the approach extensively using real experimental data, as well as synthetic data consisting of model-generated signals with real ocean noise. The results demonstrate significant improvements in model prediction accuracy, underscoring the potential of the method to enhance underwater acoustic localization in diverse, noisy, and unknown environments.

♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025

Huang-Cheng Chou, Haibin Wu, Chi-Chun Lee

Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.

comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025

♻ ☆ Speech Enhancement and Dereverberation with Diffusion-based Generative Models

Julius Richter, Simon Welker, Jean-Marie Lemercier, Bunlong Lay, Timo Gerkmann

In this work, we build upon our previous publication and use diffusion-based generative models for speech enhancement. We present a detailed overview of the diffusion process that is based on a stochastic differential equation and delve into an extensive theoretical examination of its implications. Opposed to usual conditional generation tasks, we do not start the reverse process from pure Gaussian noise but from a mixture of noisy speech and Gaussian noise. This matches our forward process which moves from clean speech to noisy speech by including a drift term. We show that this procedure enables using only 30 diffusion steps to generate high-quality clean speech estimates. By adapting the network architecture, we are able to significantly improve the speech enhancement performance, indicating that the network, rather than the formalism, was the main limitation of our original approach. In an extensive cross-dataset evaluation, we show that the improved method can compete with recent discriminative models and achieves better generalization when evaluating on a different corpus than used for training. We complement the results with an instrumental evaluation using real-world noisy recordings and a listening experiment, in which our proposed method is rated best. Examining different sampler configurations for solving the reverse process allows us to balance the performance and computational speed of the proposed method. Moreover, we show that the proposed method is also suitable for dereverberation and thus not limited to additive background noise removal. Code and audio examples are available online, see https://github.com/sp-uhh/sgmse.

comment: Proofread version

♻ ☆ Modeling nonuniform energy decay through the modal decomposition of acoustic radiance transfer (MoD-ART)

Matteo Scerbo, Sebastian J. Schlecht, Randall Ali, Lauri Savioja, Enzo De Sena

Modeling late reverberation in real-time interactive applications is a challenging task when multiple sound sources and listeners are present in the same environment. This is especially problematic when the environment is geometrically complex and/or features uneven energy absorption (e.g. coupled volumes), because in such cases the late reverberation is dependent on the sound sources' and listeners' positions, and therefore must be adapted to their movements in real time. We present a novel approach to the task, named modal decomposition of acoustic radiance transfer (MoD-ART), which can handle highly complex scenarios with efficiency. The approach is based on the geometrical acoustics method of acoustic radiance transfer, from which we extract a set of energy decay modes and their positional relationships with sources and listeners. In this paper, we describe the physical and mathematical significance of MoD-ART, highlighting its advantages and applicability to different scenarios. Through an analysis of the method's computational complexity, we show that it compares very favorably with ray-tracing. We also present simulation results showing that MoD-ART can capture multiple decay slopes and flutter echoes.

☆ Automatic Music Sample Identification with Multi-Track Contrastive Learning

Alain Riou, Joan Serrà, Yuki Mitsufuji

Sampling, the technique of reusing pieces of existing audio tracks to create new music content, is a very common practice in modern music production. In this paper, we tackle the challenging task of automatic sample identification, that is, detecting such sampled content and retrieving the material from which it originates. To do so, we adopt a self-supervised learning approach that leverages a multi-track dataset to create positive pairs of artificial mixes, and design a novel contrastive learning objective. We show that such method significantly outperforms previous state-of-the-art baselines, that is robust to various genres, and that scales well when increasing the number of noise songs in the reference database. In addition, we extensively analyze the contribution of the different components of our training pipeline and highlight, in particular, the need for high-quality separated stems for this task.

☆ ILD-VIT: A Unified Vision Transformer Architecture for Detection of Interstitial Lung Disease from Respiratory Sounds

Soubhagya Ranjan Hota, Arka Roy, Udit Satija

Interstitial lung disease (ILD) represents a group of restrictive chronic pulmonary diseases that impair oxygen acquisition by causing irreversible changes in the lungs such as fibrosis, scarring of parenchyma, etc. ILD conditions are often diagnosed by various clinical modalities such as spirometry, high-resolution lung imaging techniques, crackling respiratory sounds (RSs), etc. In this letter, we develop a novel vision transformer (VIT)-based deep learning framework namely, ILD-VIT, to detect the ILD condition using the RS recordings. The proposed framework comprises three major stages: pre-processing, mel spectrogram extraction, and classification using the proposed VIT architecture using the mel spectrogram image patches. Experimental results using the publicly available BRACETS and KAUH databases show that our proposed ILD-VIT achieves an accuracy, sensitivity, and specificity of 84.86%, 82.67%, and 86.91%, respectively, for subject-independent blind testing. The successful onboard implantation of the proposed framework on a Raspberry-pi-4 microcontroller indicates its potential as a standalone clinical system for ILD screening in a real clinical scenario.

☆ Dynamically Slimmable Speech Enhancement Network with Metric-Guided Training ICASSP2026

Haixin Zhao, Kaixuan Yang, Nilesh Madhu

To further reduce the complexity of lightweight speech enhancement models, we introduce a gating-based Dynamically Slimmable Network (DSN). The DSN comprises static and dynamic components. For architecture-independent applicability, we introduce distinct dynamic structures targeting the commonly used components, namely, grouped recurrent neural network units, multi-head attention, convolutional, and fully connected layers. A policy module adaptively governs the use of dynamic parts at a frame-wise resolution according to the input signal quality, controlling computational load. We further propose Metric-Guided Training (MGT) to explicitly guide the policy module in assessing input speech quality. Experimental results demonstrate that the DSN achieves comparable enhancement performance in instrumental metrics to the state-of-the-art lightweight baseline, while using only 73% of its computational load on average. Evaluations of dynamic component usage ratios indicate that the MGT-DSN can appropriately allocate network resources according to the severity of input signal distortion.

comment: Preprint version of a paper under review at ICASSP2026

☆ Phase Aware Ear-Conditioned Learning for Multi-Channel Binaural Speaker Separation

Ruben Johnson Robert Jeremiah, Peyman Goli, Steven van de Par

Separating competing speech in reverberant environments requires models that preserve spatial cues while maintaining separation efficiency. We present a Phase-aware Ear-conditioned speaker Separation network using eight microphones (PEASE-8) that consumes complex STFTs and directly introduces a raw-STFT input to the early decoder layer, bypassing the entire encoder pathway to improve reconstruction. The model is trained end-to-end with an SI-SDR-based objective against direct-path ear targets, jointly performing separation and dereverberation for two speakers in a fixed azimuth, eliminating the need for permutation invariant training. On spatialized two-speaker mixtures spanning anechoic, reverberant, and noisy conditions, PEASE-8 delivers strong separation and intelligibility. In reverberant environments, it achieves 12.37 dB SI-SDR, 0.87 STOI, and 1.86 PESQ at T60 = 0.6 s, while remaining competitive under anechoic conditions.

☆ Diffusion-Link: Diffusion Probabilistic Model for Bridging the Audio-Text Modality Gap ICASSP 2026

KiHyun Nam, Jongmin Choi, Hyeongkeun Lee, Jungwoo Heo, Joon Son Chung

Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link

comment: 5 pages. Submitted to IEEE ICASSP 2026

☆ Efficient Edge Test-Time Adaptation via Latent Feature Coordinate Correction

Xinyu Luo, Jie Liu, Kecheng Chen, Junyi Yang, Bo Ding, Arindam Basu, Haoliang Li

Edge devices face significant challenges due to limited computational resources and distribution shifts, making efficient and adaptable machine learning essential. Existing test-time adaptation (TTA) methods often rely on gradient-based optimization or batch processing, which are inherently unsuitable for resource-constrained edge scenarios due to their reliance on backpropagation and high computational demands. Gradient-free alternatives address these issues but often suffer from limited learning capacity, lack flexibility, or impose architectural constraints. To overcome these limitations, we propose a novel single-instance TTA method tailored for edge devices (TED), which employs forward-only coordinate optimization in the principal subspace of latent using the covariance matrix adaptation evolution strategy (CMA-ES). By updating a compact low-dimensional vector, TED not only enhances output confidence but also aligns the latent representation closer to the source latent distribution within the latent principal subspace. This is achieved without backpropagation, keeping the model parameters frozen, and enabling efficient, forgetting-free adaptation with minimal memory and computational overhead. Experiments on image classification and keyword spotting tasks across the ImageNet and Google Speech Commands series datasets demonstrate that TED achieves state-of-the-art performance while $\textit{reducing computational complexity by up to 63 times}$, offering a practical and scalable solution for real-world edge applications. Furthermore, we successfully $\textit{deployed TED on the ZYNQ-7020 platform}$, demonstrating its feasibility and effectiveness for resource-constrained edge devices in real-world deployments.

comment: Under review

☆ Unify Variables in Neural Scaling Laws for General Audio Representations via Embedding Effective Rank

Xuyao Deng, Yanjie Sun, Yong Dou, Kele Xu

Scaling laws have profoundly shaped our understanding of model performance in computer vision and natural language processing, yet their application to general audio representation learning remains underexplored. A key challenge lies in the multifactorial nature of general audio representation-representation quality is jointly influenced by variables such as audio length, embedding dimensionality, model depth, model architecture, data volume, etc., many of which are difficult to isolate or express analytically. In this work, we present a systematic study of scaling laws for general audio representations by utilizing embedding effective rank (RankMe) as a unifying metric that encapsulates the impact of diverse variables on representation quality. RankMe enables a label-free, information-theoretic quantification of audio embeddings, allowing us to examine scaling behaviors across a wide hyper-parameter space, including model size, training data volume, computational budget, architectural configurations, etc. Our empirical findings reveal a consistent power-law relationship between RankMe and representation quality, suggesting that embedding effective rank serves as a reliable proxy for assessing and predicting model performance in audio representation learning. This work not only validates the applicability of classical scaling principles to the general audio domain but also offers a theoretically grounded and empirically robust framework for guiding future model scaling strategies in audio foundation models.

☆ Delayed 1T to 2H Phase Transition Upon Electrochemical Delithiation of LiMoS2

Yerin Hong, Juhwan Lim, Jinhong Min, Nishkarsh Agarwal, Robert Hovden, Ageeth A. Bol, Yiyang Li

Molybdenum disulfide (MoS2) is a widely studied layered material for electronic, optical, and catalytic applications. It can host lithium ions between the van der Waals layers, which triggers a phase transition between the semiconducting 2H phase and metallic 1T phase. While lithium insertion triggers a phase transition to the 1T phase, the phase behavior upon electrochemical lithium removal is not resolved. In this work, we conduct single-flake electrochemical (de)lithiation of MoS2 using microelectrode arrays. Through both electrochemical voltage analysis and correlative Raman spectroscopy, we show that an electrochemically cycled and delithiated MoS2 flake initially remains in the 1T phase. However, over the course of several days, it transitions back into the thermodynamically stable 2H phase. This result resolves the phase transformation pathway upon delithiation and showcases the ability to electrochemically synthesize the metastable 1T-MoS2 phase.

☆ Perceptual Compensation of Ambisonics Recordings for Reproduction in Room

Ali Fallah, Shun Nakamura, Steven van de Par

Ambisonics is a method for capturing and rendering a sound field accurately, assuming that the acoustics of the playback room does not significantly influence the sound field. However, in practice, the acoustics of the playback room may lead to a noticeable degradation in sound quality. We propose a recording and rendering method based on Ambisonics that utilizes a perceptually-motivated approach to compensate for the reverberation of the playback room. The recorded direct and reverberant sound field components in the spherical harmonics (SHs) domain are spectrally and spatially compensated to preserve the relevant auditory cues including the direction of arrival of the direct sound, the spectral energy of the direct and reverberant sound components, and the Interaural Coherence (IC) across each auditory band. In contrast to the conventional Ambisonics, a flexible number of Ambisonics channels can be used for audio rendering. Listening test results show that the proposed method provides a perceptually accurate rendering of the originally recorded sound field, outperforming both conventional Ambisonics without compensation and even ideal Ambisonics rendering in a simulated anechoic room. Additionally, subjective evaluations of listeners seated at the center of the loudspeaker array demonstrate that the method remains robust to head rotation and minor displacements.

comment: The manuscript was submitted to the JASA and is under review

☆ Gelina: Unified Speech and Gesture Synthesis via Interleaved Token Prediction

Téo Guichoux, Théodor Lemerle, Shivam Mehta, Jonas Beskow, Gustave Eje Henter, Laure Soulier, Catherine Pelachaud, Nicolas Obin

Human communication is multimodal, with speech and gestures tightly coupled, yet most computational methods for generating speech and gestures synthesize them sequentially, weakening synchrony and prosody alignment. We introduce Gelina, a unified framework that jointly synthesizes speech and co-speech gestures from text using interleaved token sequences in a discrete autoregressive backbone, with modality-specific decoders. Gelina supports multi-speaker and multi-style cloning and enables gesture-only synthesis from speech inputs. Subjective and objective evaluations demonstrate competitive speech quality and improved gesture generation over unimodal baselines.

comment: 5 pages

♻ ☆ MGE-LDM: Joint Latent Diffusion for Simultaneous Music Generation and Source Extraction NeurIPS 2025

Yunkee Chae, Kyogu Lee

We present MGE-LDM, a unified latent diffusion framework for simultaneous music generation, source imputation, and query-driven source separation. Unlike prior approaches constrained to fixed instrument classes, MGE-LDM learns a joint distribution over full mixtures, submixtures, and individual stems within a single compact latent diffusion model. At inference, MGE-LDM enables (1) complete mixture generation, (2) partial generation (i.e., source imputation), and (3) text-conditioned extraction of arbitrary sources. By formulating both separation and imputation as conditional inpainting tasks in the latent space, our approach supports flexible, class-agnostic manipulation of arbitrary instrument sources. Notably, MGE-LDM can be trained jointly across heterogeneous multi-track datasets (e.g., Slakh2100, MUSDB18, MoisesDB) without relying on predefined instrument categories. Audio samples are available at our project page: https://yoongi43.github.io/MGELDM_Samples/.

comment: Accepted by NeurIPS 2025

♻ ☆ GRAM: Spatial general-purpose audio representation models for real-world applications

Goksenin Yuksel, Marcel van Gerven, Kiki van der Heijden

Although audio foundations models have seen great progress on a wide variety of tasks, their application in real-world acoustic environments with reverberation and noise has been less successful. Moreover, as audio foundation models are typically trained on dry, single-channel audio clips, the inherent spatial nature of real-world sound scenes is overlooked and tasks involving sound localization ruled out. To address these limitations, we propose GRAM: a General-purpose Real-world Audio Model utilizing a multi-channel masked auto-encoder approach to efficiently learn spatial audio representations from high-quality simulated real-world scenes. To evaluate the performance of GRAM and other audio foundation models in real-world sound scenes, we release Nat-HEAR: A naturalistic version of the HEAR benchmark suite comprising a simulated real-world version, as well as two new sound localization tasks. We show that the performance of GRAM surpasses all state-of-the-art self-supervised audio foundation models and speech models on both HEAR and Nat-HEAR, while using only a fraction of the training data. GRAM also showcases state-of-the-art localization performance, surpassing even supervised sound localization approaches, and can be flexibly applied either to a two-channel, binaural sound format or a four-channel, Ambisonics format. Validating GRAM's performance on real-world sound recordings demonstrates robust transfer to real-world scenes. Taken together, GRAM presents a significant advancement towards robust, spatial audio foundation models for real-world applications.

comment: Still under review

♻ ☆ WavJEPA: Semantic learning unlocks robust audio foundation models for raw waveforms

Goksenin Yuksel, Pierre Guetschel, Michael Tangermann, Marcel van Gerven, Kiki van der Heijden

Learning audio representations from raw waveforms overcomes key limitations of spectrogram-based audio representation learning, such as the long latency of spectrogram computation and the loss of phase information. Yet, while self-supervised speech representation learning from raw waveforms has been remarkably successful, these approaches have not achieved similar feats for general-purpose audio representation learning from waveforms. Here, we propose WavJEPA, a waveform-based version of the Joint-Embedding Predictive Architecture. WavJEPA leverages high-level semantic representation learning to tackle the shortcomings of representation learning at the speech unit or token level. We show that this approach substantially outperforms state-of-the-art time-domain audio foundation models across a wide variety of downstream benchmark tasks, while requiring considerably fewer computational resources. Additionally, to overcome the performance drop that time-domain models typically exhibit in noisy and reverberant real-world acoustic environments, we present WavJEPA-Nat. WavJEPA-Nat is a multi-channel extension of the WavJEPA architecture trained on simulated naturalistic scenes. We find that WavJEPA-Nat is highly robust to reverberation and noise. These results highlight the feasibility and computational efficiency of general-purpose audio representation learning from raw waveforms, showcasing the potential for low-latency, robust time-domain audio foundation models for real-world applications.

comment: Still under review

♻ ☆ Enhancing Noise Robustness for Neural Speech Codecs through Resource-Efficient Progressive Quantization Perturbation Simulation

Rui-Chen Zheng, Yang Ai, Hui-Peng Du, Li-Rong Dai

Noise robustness remains a critical challenge for deploying neural speech codecs in real-world acoustic scenarios where background noise is often inevitable. A key observation we make is that even slight input noise perturbations can cause unintended shifts in quantized codewords, thereby degrading the quality of reconstructed speech. Motivated by this finding, we propose a novel and resource-efficient training strategy to enhance the noise robustness of speech codecs by simulating such perturbations directly at the quantization level. Our approach introduces two core mechanisms: (1) a distance-weighted probabilistic top-K sampling strategy that replaces the conventional deterministic nearest-neighbor selection in residual vector quantization (RVQ); and (2) a progressive training scheme that introduces perturbations from the last to the first quantizer in a controlled manner. Crucially, our method is trained exclusively on clean speech, eliminating the need for any paired noisy-clean data. Experiments on two advanced neural speech codecs, Encodec and WavTokenizer, demonstrate that the proposed strategy substantially improves robustness under noisy conditions-for example, boosting UTMOS from 3.475 to 3.586 at 15 dB SNR on Encodec-while also enhancing coding quality for clean speech.

♻ ☆ Discrete-Time Diffusion-Like Models for Speech Synthesis

Xiaozhou Tan, Minghui Zhao, Anton Ragni

Diffusion models have attracted a lot of attention in recent years. These models view speech generation as a continuous-time process. For efficient training, this process is typically restricted to additive Gaussian noising, which is limiting. For inference, the time is typically discretized, leading to the mismatch between continuous training and discrete sampling conditions. Recently proposed discrete-time processes, on the other hand, usually do not have these limitations, may require substantially fewer inference steps, and are fully consistent between training/inference conditions. This paper explores some diffusion-like discrete-time processes and proposes some new variants. These include processes applying additive Gaussian noise, multiplicative Gaussian noise, blurring noise and a mixture of blurring and Gaussian noises. The experimental results suggest that discrete-time processes offer comparable subjective and objective speech quality to their widely popular continuous counterpart, with more efficient and consistent training and inference schemas.

♻ ☆ Joint Source-Environment Adaptation of Data-Driven Underwater Acoustic Source Ranging Based on Model Uncertainty

Dariush Kari, Hari Vishnu, Andrew C. Singer

Adapting pre-trained deep learning models to new and unknown environments remains a major challenge in underwater acoustic localization. We show that although the performance of pre-trained models suffers from mismatch between the training and test data, they generally exhibit a higher uncertainty in environments where there is more mismatch. Additionally, in the presence of environmental mismatch, spurious peaks can appear in the output of classification-based localization approaches, which inspires us to define and use a method to quantify the "implied uncertainty" based on the number of model output peaks. Leveraging this notion of implied uncertainty, we partition the test samples into sets with more certain and less certain samples, and implement a method to adapt the model to new environments by using the certain samples to improve the labeling for uncertain samples, which helps to adapt the model. Thus, using this efficient method for model uncertainty quantification, we showcase an innovative approach to adapt a pre-trained model to unseen underwater environments at test time. This eliminates the need for labeled data from the target environment or the original training data. This adaptation is enhanced by integrating an independent estimate based on the received signal energy. We validate the approach extensively using real experimental data, as well as synthetic data consisting of model-generated signals with real ocean noise. The results demonstrate significant improvements in model prediction accuracy, underscoring the potential of the method to enhance underwater acoustic localization in diverse, noisy, and unknown environments.

♻ ☆ Stimulus Modality Matters: Impact of Perceptual Evaluations from Different Modalities on Speech Emotion Recognition System Performance ICASSP 2025

Huang-Cheng Chou, Haibin Wu, Chi-Chun Lee

Speech Emotion Recognition (SER) systems rely on speech input and emotional labels annotated by humans. However, various emotion databases collect perceptional evaluations in different ways. For instance, the IEMOCAP dataset uses video clips with sounds for annotators to provide their emotional perceptions. However, the most significant English emotion dataset, the MSP-PODCAST, only provides speech for raters to choose the emotional ratings. Nevertheless, using speech as input is the standard approach to training SER systems. Therefore, the open question is the emotional labels elicited by which scenarios are the most effective for training SER systems. We comprehensively compare the effectiveness of SER systems trained with labels elicited by different modality stimuli and evaluate the SER systems on various testing conditions. Also, we introduce an all-inclusive label that combines all labels elicited by various modalities. We show that using labels elicited by voice-only stimuli for training yields better performance on the test set, whereas labels elicited by voice-only stimuli.

comment: 5 pages, 2 figures, 4 tables, acceptance for ICASSP 2025

♻ ☆ Speech Enhancement and Dereverberation with Diffusion-based Generative Models

Julius Richter, Simon Welker, Jean-Marie Lemercier, Bunlong Lay, Timo Gerkmann

In this work, we build upon our previous publication and use diffusion-based generative models for speech enhancement. We present a detailed overview of the diffusion process that is based on a stochastic differential equation and delve into an extensive theoretical examination of its implications. Opposed to usual conditional generation tasks, we do not start the reverse process from pure Gaussian noise but from a mixture of noisy speech and Gaussian noise. This matches our forward process which moves from clean speech to noisy speech by including a drift term. We show that this procedure enables using only 30 diffusion steps to generate high-quality clean speech estimates. By adapting the network architecture, we are able to significantly improve the speech enhancement performance, indicating that the network, rather than the formalism, was the main limitation of our original approach. In an extensive cross-dataset evaluation, we show that the improved method can compete with recent discriminative models and achieves better generalization when evaluating on a different corpus than used for training. We complement the results with an instrumental evaluation using real-world noisy recordings and a listening experiment, in which our proposed method is rated best. Examining different sampler configurations for solving the reverse process allows us to balance the performance and computational speed of the proposed method. Moreover, we show that the proposed method is also suitable for dereverberation and thus not limited to additive background noise removal. Code and audio examples are available online, see https://github.com/sp-uhh/sgmse.

comment: Proofread version

♻ ☆ Modeling nonuniform energy decay through the modal decomposition of acoustic radiance transfer (MoD-ART)

Matteo Scerbo, Sebastian J. Schlecht, Randall Ali, Lauri Savioja, Enzo De Sena

Modeling late reverberation in real-time interactive applications is a challenging task when multiple sound sources and listeners are present in the same environment. This is especially problematic when the environment is geometrically complex and/or features uneven energy absorption (e.g. coupled volumes), because in such cases the late reverberation is dependent on the sound sources' and listeners' positions, and therefore must be adapted to their movements in real time. We present a novel approach to the task, named modal decomposition of acoustic radiance transfer (MoD-ART), which can handle highly complex scenarios with efficiency. The approach is based on the geometrical acoustics method of acoustic radiance transfer, from which we extract a set of energy decay modes and their positional relationships with sources and listeners. In this paper, we describe the physical and mathematical significance of MoD-ART, highlighting its advantages and applicability to different scenarios. Through an analysis of the method's computational complexity, we show that it compares very favorably with ray-tracing. We also present simulation results showing that MoD-ART can capture multiple decay slopes and flutter echoes.

☆ FAC-FACodec: Controllable Zero-Shot Foreign Accent Conversion with Factorized Speech Codec

Yurii Halychanskyi, Cameron Churchwell, Yutong Wen, Volodymyr Kindratenko

Previous accent conversion (AC) methods, including foreign accent conversion (FAC), lack explicit control over the degree of modification. Because accent modification can alter the perceived speaker identity, balancing conversion strength and identity preservation is crucial. We present an AC framework that provides an explicit, user-controllable parameter for accent modification. The method targets pronunciation while preserving suprasegmental cues such as intonation and phoneme durations. Results show performance comparable to recent AC systems, stronger preservation of speaker identity, and unique support for controllable accent conversion.

comment: 5 pages, 2 figures

☆ ParsVoice: A Large-Scale Multi-Speaker Persian Speech Corpus for Text-to-Speech Synthesis

Mohammad Javad Ranjbar Kalahroodi, Heshaam Faili, Azadeh Shakery

Persian Language, despite being spoken by over 100 million people worldwide, remains severely underrepresented in high-quality speech corpora, particularly for text-to-speech (TTS) synthesis applications. Existing Persian speech datasets are typically smaller than their English counterparts, which creates a key limitation for developing Persian speech technologies. We address this gap by introducing ParsVoice, the largest Persian speech corpus designed specifically for TTS applications. We created an automated pipeline that transforms raw audiobook content into TTS-ready data, incorporating components such as a BERT-based sentence completion detector, a binary search boundary optimization method for precise audio-text alignment, and multi-dimensional quality assessment frameworks tailored to Persian. The pipeline processes 2,000 audiobooks, yielding 3,526 hours of clean speech, which was further filtered into a 1,804-hour high-quality subset suitable for TTS, featuring more than 470 speakers. ParsVoice is the largest high-quality Persian speech dataset, offering speaker diversity and audio quality comparable to major English corpora. The complete dataset has been made publicly available to accelerate the development of Persian speech technologies and to serve as a template for other low-resource languages. The ParsVoice dataset is publicly available at ParsVoice (https://huggingface.co/datasets/MohammadJRanjbar/ParsVoice).

☆ Dual Data Scaling for Robust Two-Stage User-Defined Keyword Spotting

Zhiqi Ai, Han Cheng, Yuxin Wang, Shiyi Mu, Shugong Xu, Yongjin Zhou

In this paper, we propose DS-KWS, a two-stage framework for robust user-defined keyword spotting. It combines a CTC-based method with a streaming phoneme search module to locate candidate segments, followed by a QbyT-based method with a phoneme matcher module for verification at both the phoneme and utterance levels. To further improve performance, we introduce a dual data scaling strategy: (1) expanding the ASR corpus from 460 to 1,460 hours to strengthen the acoustic model; and (2) leveraging over 155k anchor classes to train the phoneme matcher, significantly enhancing the distinction of confusable words. Experiments on LibriPhrase show that DS-KWS significantly outperforms existing methods, achieving 6.13\% EER and 97.85\% AUC on the Hard subset. On Hey-Snips, it achieves zero-shot performance comparable to full-shot trained models, reaching 99.13\% recall at one false alarm per hour.

comment: 5 pages, 3 figures

☆ Proficiency-Aware Adaptation and Data Augmentation for Robust L2 ASR ICASSP 2026

Ling Sun, Charlotte Zhu, Shuju Shi

General-purpose ASR underperforms for atypical speakers, such as L2 learners, reinforcing bias and limiting use in education and accessibility. Using the CEFR-graded Speak and Improve corpus, we show that naive fine-tuning of Whisper reduces average WER but simultaneously widens disparities and disproportionately harms lower-level learners. To address this, we propose two strategies: (i) proficiency-aware multitask learning, jointly optimizing ASR with proficiency classification, and (ii) targeted augmentation, applying spectrogram masking to low-proficiency speech to counter imbalance. These approaches reduce WER by up to 29.4 percent (relative) and insertion/deletion errors by as much as 58.6 percent (relative). Crucially, despite the severe imbalance of the dataset reflecting real-world distributions, both strategies consistently narrow proficiency gaps, advancing equitable ASR for L2 learners.

comment: Submitted to ICASSP 2026

☆ SS-DPPN: A self-supervised dual-path foundation model for the generalizable cardiac audio representation

Ummy Maria Muna, Md Mehedi Hasan Shawon, Md Jobayer, Sumaiya Akter, Md Rakibul Hasan, Md. Golam Rabiul Alam

The automated analysis of phonocardiograms is vital for the early diagnosis of cardiovascular disease, yet supervised deep learning is often constrained by the scarcity of expert-annotated data. In this paper, we propose the Self-Supervised Dual-Path Prototypical Network (SS-DPPN), a foundation model for cardiac audio representation and classification from unlabeled data. The framework introduces a dual-path contrastive learning based architecture that simultaneously processes 1D waveforms and 2D spectrograms using a novel hybrid loss. For the downstream task, a metric-learning approach using a Prototypical Network was used that enhances sensitivity and produces well-calibrated and trustworthy predictions. SS-DPPN achieves state-of-the-art performance on four cardiac audio benchmarks. The framework demonstrates exceptional data efficiency with a fully supervised model on three-fold reduction in labeled data. Finally, the learned representations generalize successfully across lung sound classification and heart rate estimation. Our experiments and findings validate SS-DPPN as a robust, reliable, and scalable foundation model for physiological signals.

☆ LSZone: A Lightweight Spatial Information Modeling Architecture for Real-time In-car Multi-zone Speech Separation ICASSP 2026

Jun Chen, Shichao Hu, Jiuxin Lin, Wenjie Li, Zihan Zhang, Xingchen Li, JinJiang Liu, Longshuai Xiao, Chao Weng, Lei Xie, Zhiyong Wu

In-car multi-zone speech separation, which captures voices from different speech zones, plays a crucial role in human-vehicle interaction. Although previous SpatialNet has achieved notable results, its high computational cost still hinders real-time applications in vehicles. To this end, this paper proposes LSZone, a lightweight spatial information modeling architecture for real-time in-car multi-zone speech separation. We design a spatial information extraction-compression (SpaIEC) module that combines Mel spectrogram and Interaural Phase Difference (IPD) to reduce computational burden while maintaining performance. Additionally, to efficiently model spatial information, we introduce an extremely lightweight Conv-GRU crossband-narrowband processing (CNP) module. Experimental results demonstrate that LSZone, with a complexity of 0.56G MACs and a real-time factor (RTF) of 0.37, delivers impressive performance in complex noise and multi-speaker scenarios.

comment: submitted to ICASSP 2026

☆ A Machine Learning Approach for MIDI to Guitar Tablature Conversion

Maximos Kaliakatsos-Papakostas, Gregoris Bastas, Dimos Makris, Dorien Herremans, Vassilis Katsouros, Petros Maragos

Guitar tablature transcription consists in deducing the string and the fret number on which each note should be played to reproduce the actual musical part. This assignment should lead to playable string-fret combinations throughout the entire track and, in general, preserve parsimonious motion between successive combinations. Throughout the history of guitar playing, specific chord fingerings have been developed across different musical styles that facilitate common idiomatic voicing combinations and motion between them. This paper presents a method for assigning guitar tablature notation to a given MIDI-based musical part (possibly consisting of multiple polyphonic tracks), i.e. no information about guitar-idiomatic expressional characteristics is involved (e.g. bending etc.) The current strategy is based on machine learning and requires a basic assumption about how much fingers can stretch on a fretboard; only standard 6-string guitar tuning is examined. The proposed method also examines the transcription of music pieces that was not meant to be played or could not possibly be played by a guitar (e.g. potentially a symphonic orchestra part), employing a rudimentary method for augmenting musical information and training/testing the system with artificial data. The results present interesting aspects about what the system can achieve when trained on the initial and augmented dataset, showing that the training with augmented data improves the performance even in simple, e.g. monophonic, cases. Results also indicate weaknesses and lead to useful conclusions about possible improvements.

comment: Proceedings of the 19th Sound and Music Computing Conference, June 5-12th, 2022, Saint-\'Etienne (France)

☆ MARS-Sep: Multimodal-Aligned Reinforced Sound Separation

Zihan Zhang, Xize Cheng, Zhennan Jiang, Dongjie Fu, Jingyuan Chen, Zhou Zhao, Tao Jin

Universal sound separation faces a fundamental misalignment: models optimized for low-level signal metrics often produce semantically contaminated outputs, failing to suppress perceptually salient interference from acoustically similar sources. To bridge this gap, we introduce MARS-Sep, a reinforcement learning framework that reformulates separation as decision making. Instead of simply regressing ground-truth masks, MARS-Sep learns a factorized Beta mask policy that is optimized by a clipped trust-region surrogate with entropy regularization and group-relative advantage normalization. Concretely, we sample masks from a frozen old policy, reconstruct waveforms, and update the current policy using clipped importance ratios-yielding substantially more stable and sample-efficient learning. Multimodal rewards, derived from an audio-text-vision encoder, directly incentivize semantic consistency with query prompts. We further propose a progressive alignment scheme to fine-tune this encoder, boosting its cross-modal discriminability and improving reward faithfulness. Extensive experiments on multiple benchmarks demonstrate consistent gains in Text-, Audio-, and Image-Queried separation, with notable improvements in signal metrics and semantic quality. Our code is available at https://anonymous.4open.science/r/MARS-Sep. Sound separation samples are available at https://mars-sep.github.io/.

☆ Knowledge-Decoupled Functionally Invariant Path with Synthetic Personal Data for Personalized ASR

Yue Gu, Zhihao Du, Ying Shi, Jiqing Han, Yongjun He

Fine-tuning generic ASR models with large-scale synthetic personal data can enhance the personalization of ASR models, but it introduces challenges in adapting to synthetic personal data without forgetting real knowledge, and in adapting to personal data without forgetting generic knowledge. Considering that the functionally invariant path (FIP) framework enables model adaptation while preserving prior knowledge, in this letter, we introduce FIP into synthetic-data-augmented personalized ASR models. However, the model still struggles to balance the learning of synthetic, personalized, and generic knowledge when applying FIP to train the model on all three types of data simultaneously. To decouple this learning process and further address the above two challenges, we integrate a gated parameter-isolation strategy into FIP and propose a knowledge-decoupled functionally invariant path (KDFIP) framework, which stores generic and personalized knowledge in separate modules and applies FIP to them sequentially. Specifically, KDFIP adapts the personalized module to synthetic and real personal data and the generic module to generic data. Both modules are updated along personalization-invariant paths, and their outputs are dynamically fused through a gating mechanism. With augmented synthetic data, KDFIP achieves a 29.38% relative character error rate reduction on target speakers and maintains comparable generalization performance to the unadapted ASR baseline.

comment: Accepted for publication in IEEE Signal Processing Letters, 2025

☆ MRSAudio: A Large-Scale Multimodal Recorded Spatial Audio Dataset with Refined Annotations

Wenxiang Guo, Changhao Pan, Zhiyuan Zhu, Xintong Hu, Yu Zhang, Li Tang, Rui Yang, Han Wang, Zongbao Zhang, Yuhan Wang, Yixuan Chen, Hankun Xu, Ke Xu, Pengfei Fan, Zhetao Chen, Yanhao Yu, Qiange Huang, Fei Wu, Zhou Zhao

Humans rely on multisensory integration to perceive spatial environments, where auditory cues enable sound source localization in three-dimensional space. Despite the critical role of spatial audio in immersive technologies such as VR/AR, most existing multimodal datasets provide only monaural audio, which limits the development of spatial audio generation and understanding. To address these challenges, we introduce MRSAudio, a large-scale multimodal spatial audio dataset designed to advance research in spatial audio understanding and generation. MRSAudio spans four distinct components: MRSLife, MRSSpeech, MRSMusic, and MRSSing, covering diverse real-world scenarios. The dataset includes synchronized binaural and ambisonic audio, exocentric and egocentric video, motion trajectories, and fine-grained annotations such as transcripts, phoneme boundaries, lyrics, scores, and prompts. To demonstrate the utility and versatility of MRSAudio, we establish five foundational tasks: audio spatialization, and spatial text to speech, spatial singing voice synthesis, spatial music generation and sound event localization and detection. Results show that MRSAudio enables high-quality spatial modeling and supports a broad range of spatial audio research. Demos and dataset access are available at https://mrsaudio.github.io.

comment: 24 pages

♻ ☆ Data Standards in Audiology: A Mixed-Methods Exploration of Community Perspectives and Implementation Considerations

Charlotte Vercammen, Antje Heinrich, Christophe Lesimple, Alessia Paglialonga, Jan-Willem A. Wasmann, Mareike Buhl

Objective: This study addresses conceptual issues around data standardisation in audiology, and outlines steps toward achieving it. It reports a survey of the computational audiology community on their current understanding, needs, and preferences concerning data standards. Based on survey findings and a panel discussion, recommendations are made concerning moving forward with standardisation in audiology. Design: Mixed-methods: 1) review of existing standardisation efforts; 2) a survey of the computational audiology community; 3) expert panel discussion in a dedicated session at the 2024 Virtual Conference of Computational Audiology. Sample: Survey: 82 members of the global community; Panel discussion: five experts. Results: A prerequisite for any global audiology database are agreed data standards. Although many are familiar with the general idea, few know of existing initiatives, or have actively participated in them. Ninety percent of respondents expressed willingness to follow or contribute to standardisation efforts. The panel discussed relevant initiatives (e.g. OMOP, openEHR, NOAH) and explored both challenges (around harmonisation) and opportunities (alignment with other medical fields and conversion among approaches). Conclusions: Combining conceptual discussion with stakeholder views, the study offers guidance for implementing interoperable data standards in audiology. It highlights community support, key issues to address, and suggests paths for future work.

☆ Bhasha-Rupantarika: Algorithm-Hardware Co-design approach for Multilingual Neural Machine Translation

Mukul Lokhande, Tanushree Dewangan, Mohd Sharik Mansoori, Tejas Chaudhari, Akarsh J., Damayanti Lokhande, Adam Teman, Santosh Kumar Vishvakarma

This paper introduces Bhasha-Rupantarika, a light and efficient multilingual translation system tailored through algorithm-hardware codesign for resource-limited settings. The method investigates model deployment at sub-octet precision levels (FP8, INT8, INT4, and FP4), with experimental results indicating a 4.1x reduction in model size (FP4) and a 4.2x speedup in inference speed, which correlates with an increased throughput of 66 tokens/s (improvement by 4.8x). This underscores the importance of ultra-low precision quantization for real-time deployment in IoT devices using FPGA accelerators, achieving performance on par with expectations. Our evaluation covers bidirectional translation between Indian and international languages, showcasing its adaptability in low-resource linguistic contexts. The FPGA deployment demonstrated a 1.96x reduction in LUTs and a 1.65x decrease in FFs, resulting in a 2.2x enhancement in throughput compared to OPU and a 4.6x enhancement compared to HPTA. Overall, the evaluation provides a viable solution based on quantisation-aware translation along with hardware efficiency suitable for deployable multilingual AI systems. The entire codes [https://github.com/mukullokhande99/Bhasha-Rupantarika/] and dataset for reproducibility are publicly available, facilitating rapid integration and further development by researchers.

♻ ☆ Data Standards in Audiology: A Mixed-Methods Exploration of Community Perspectives and Implementation Considerations

Charlotte Vercammen, Antje Heinrich, Christophe Lesimple, Alessia Paglialonga, Jan-Willem A. Wasmann, Mareike Buhl

Objective: This study addresses conceptual issues around data standardisation in audiology, and outlines steps toward achieving it. It reports a survey of the computational audiology community on their current understanding, needs, and preferences concerning data standards. Based on survey findings and a panel discussion, recommendations are made concerning moving forward with standardisation in audiology. Design: Mixed-methods: 1) review of existing standardisation efforts; 2) a survey of the computational audiology community; 3) expert panel discussion in a dedicated session at the 2024 Virtual Conference of Computational Audiology. Sample: Survey: 82 members of the global community; Panel discussion: five experts. Results: A prerequisite for any global audiology database are agreed data standards. Although many are familiar with the general idea, few know of existing initiatives, or have actively participated in them. Ninety percent of respondents expressed willingness to follow or contribute to standardisation efforts. The panel discussed relevant initiatives (e.g. OMOP, openEHR, NOAH) and explored both challenges (around harmonisation) and opportunities (alignment with other medical fields and conversion among approaches). Conclusions: Combining conceptual discussion with stakeholder views, the study offers guidance for implementing interoperable data standards in audiology. It highlights community support, key issues to address, and suggests paths for future work.

☆ ProGress: Structured Music Generation via Graph Diffusion and Hierarchical Music Analysis

Stephen Ni-Hahn, Chao Péter Yang, Mingchen Ma, Cynthia Rudin, Simon Mak, Yue Jiang

Artificial Intelligence (AI) for music generation is undergoing rapid developments, with recent symbolic models leveraging sophisticated deep learning and diffusion model algorithms. One drawback with existing models is that they lack structural cohesion, particularly on harmonic-melodic structure. Furthermore, such existing models are largely "black-box" in nature and are not musically interpretable. This paper addresses these limitations via a novel generative music framework that incorporates concepts of Schenkerian analysis (SchA) in concert with a diffusion modeling framework. This framework, which we call ProGress (Prolongation-enhanced DiGress), adapts state-of-the-art deep models for discrete diffusion (in particular, the DiGress model of Vignac et al., 2023) for interpretable and structured music generation. Concretely, our contributions include 1) novel adaptations of the DiGress model for music generation, 2) a novel SchA-inspired phrase fusion methodology, and 3) a framework allowing users to control various aspects of the generation process to create coherent musical compositions. Results from human experiments suggest superior performance to existing state-of-the-art methods.

☆ Peransformer: Improving Low-informed Expressive Performance Rendering with Score-aware Discriminator SC 2025

Xian He, Wei Zeng, Ye Wang

Highly-informed Expressive Performance Rendering (EPR) systems transform music scores with rich musical annotations into human-like expressive performance MIDI files. While these systems have achieved promising results, the availability of detailed music scores is limited compared to MIDI files and are less flexible to work with using a digital audio workstation (DAW). Recent advancements in low-informed EPR systems offer a more accessible alternative by directly utilizing score-derived MIDI as input, but these systems often exhibit suboptimal performance. Meanwhile, existing works are evaluated with diverse automatic metrics and data formats, hindering direct objective comparisons between EPR systems. In this study, we introduce Peransformer, a transformer-based low-informed EPR system designed to bridge the gap between low-informed and highly-informed EPR systems. Our approach incorporates a score-aware discriminator that leverages the underlying score-derived MIDI files and is trained on a score-to-performance paired, note-to-note aligned MIDI dataset. Experimental results demonstrate that Peransformer achieves state-of-the-art performance among low-informed systems, as validated by subjective evaluations. Furthermore, we extend existing automatic evaluation metrics for EPR systems and introduce generalized EPR metrics (GEM), enabling more direct, accurate, and reliable comparisons across EPR systems.

comment: 6 pages, 3 figures, accepted by APSIPA ASC 2025

☆ Chord Colourizer: A Near Real-Time System for Visualizing Musical Key

Paul Haimes

This paper introduces Chord Colourizer, a near real-time system that detects the musical key of an audio signal and visually represents it through a novel graphical user interface (GUI). The system assigns colours to musical notes based on Isaac Newton's original colour wheel, preserving historical links between pitch and hue, and also integrates an Arduino-controlled LED display using 3D-printed star-shaped diffusers to offer a physical ambient media representation. The method employs Constant-Q Transform (CQT) chroma features for chord estimation and visualization, followed by threshold-based filtering and tonal enhancement to isolate the root, third, and fifth. A confidence score is computed for each detection to ensure reliability, and only chords with moderate to very strong certainty are visualized. The graphical interface dynamically updates a colour-coded keyboard layout, while the LED display provides the same colour information via spatial feedback. This multi-modal system enhances user interaction with harmonic content, offering innovative possibilities for education and artistic performance. Limitations include slight latency and the inability to detect extended chords, which future development will aim to address through refined filtering, adaptive thresholds, and support for more complex harmonies such as sevenths and augmented chords. Future work will also explore integration with alternative visualization styles, and the comparison of audio analysis libraries to improve detection speed and precision. Plans also include formal user testing to evaluate perception, usability, and cross-cultural interpretations of colour-pitch mappings.

comment: Author copy. This paper is in press for presentation at ADADA 2025. Please cite as: Haimes, P. (in press). Chord Colourizer: A near real-time system for visualizing musical key. In Proceedings of the 23rd International Conference of Asia Digital Art and Design Association (ADADA)

☆ Matchmaker: An Open-source Library for Real-time Piano Score Following and Systematic Evaluation

Jiyun Park, Carlos Cancino-Chacón, Suhit Chiruthapudi, Juhan Nam

Real-time music alignment, also known as score following, is a fundamental MIR task with a long history and is essential for many interactive applications. Despite its importance, there has not been a unified open framework for comparing models, largely due to the inherent complexity of real-time processing and the language- or system-dependent implementations. In addition, low compatibility with the existing MIR environment has made it difficult to develop benchmarks using large datasets available in recent years. While new studies based on established methods (e.g., dynamic programming, probabilistic models) have emerged, most evaluations compare models only within the same family or on small sets of test data. This paper introduces Matchmaker, an open-source Python library for real-time music alignment that is easy to use and compatible with modern MIR libraries. Using this, we systematically compare methods along two dimensions: music representations and alignment methods. We evaluated our approach on a large test set of solo piano music from the (n)ASAP, Batik, and Vienna4x22 datasets with a comprehensive set of metrics to ensure robust assessment. Our work aims to establish a benchmark framework for score-following research while providing a practical tool that developers can easily integrate into their applications.

comment: In Proceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR), 2025

☆ Improving Speech Emotion Recognition with Mutual Information Regularized Generative Model

Chung-Soo Ahn, Rajib Rana, Sunil Sivadas, Carlos Busso, Jagath C. Rajapakse

Although speech emotion recognition (SER) research has been advanced, thanks to deep learning methods, it still suffers from obtaining inputs from large quality-labelled training data. Data augmentation methods have been attempted to mitigate this issue, generative models have shown success among them recently. We propose a data augmentation framework that is aided by cross-modal information transfer and mutual information regularization. Mutual information based metric can serve as an indicator for the quality. Furthermore, we expand this data augmentation scope to multimodal inputs, thanks to mutual information ensureing dependency between modalities. Our framework was tested on three benchmark datasets: IEMOCAP, MSP-IMPROV and MSP-Podcast. The implementation was designed to generate input features that are fed into last layer for emotion classification. Our framework improved the performance of emotion prediction against existing works. Also, we discovered that our framework is able to generate new inputs without any cross-modal information.

☆ MTP-S2UT: Enhancing Speech-to-Speech Translation Quality with Multi-token Prediction

Jianjin Wang, Runsong Zhao, Xiaoqian Liu, Yuan Ge, Ziqiang Xu, Tong Xiao, Shengxiang Gao, Zhengtao Yu, Jingbo Zhu

Current direct speech-to-speech translation methods predominantly employ speech tokens as intermediate representations. However, a single speech token is not dense in semantics, so we generally need multiple tokens to express a complete semantic unit. To address this limitation, we introduce multi-token prediction (MTP) loss into speech-to-unit translation (S2UT) models, enabling models to predict multiple subsequent tokens at each position, thereby capturing more complete semantics and enhancing information density per position. Initial MTP implementations apply the loss at the final layer, which improves output representation but initiates information enrichment too late. We hypothesize that advancing the information enrichment process to intermediate layers can achieve earlier and more effective enhancement of hidden representation. Consequently, we propose MTP-S2UT loss, applying MTP loss to hidden representation where CTC loss is computed. Experiments demonstrate that all MTP loss variants consistently improve the quality of S2UT translation, with MTP-S2UT achieving the best performance.

comment: Copyright 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works

☆ Universal Discrete-Domain Speech Enhancement

Fei Liu, Yang Ai, Ye-Xin Lu, Rui-Chen Zheng, Hui-Peng Du, Zhen-Hua Ling

In real-world scenarios, speech signals are inevitably corrupted by various types of interference, making speech enhancement (SE) a critical task for robust speech processing. However, most existing SE methods only handle a limited range of distortions, such as additive noise, reverberation, or band limitation, while the study of SE under multiple simultaneous distortions remains limited. This gap affects the generalization and practical usability of SE methods in real-world environments.To address this gap, this paper proposes a novel Universal Discrete-domain SE model called UDSE.Unlike regression-based SE models that directly predict clean speech waveform or continuous features, UDSE redefines SE as a discrete-domain classification task, instead predicting the clean discrete tokens quantized by the residual vector quantizer (RVQ) of a pre-trained neural speech codec.Specifically, UDSE first extracts global features from the degraded speech. Guided by these global features, the clean token prediction for each VQ follows the rules of RVQ, where the prediction of each VQ relies on the results of the preceding ones. Finally, the predicted clean tokens from all VQs are decoded to reconstruct the clean speech waveform. During training, the UDSE model employs a teacher-forcing strategy, and is optimized with cross-entropy loss. Experimental results confirm that the proposed UDSE model can effectively enhance speech degraded by various conventional and unconventional distortions, e.g., additive noise, reverberation, band limitation, clipping, phase distortion, and compression distortion, as well as their combinations. These results demonstrate the superior universality and practicality of UDSE compared to advanced regression-based SE methods.

☆ Automatic Speech Recognition in the Modern Era: Architectures, Training, and Evaluation

Md. Nayeem, Md Shamse Tabrej, Kabbojit Jit Deb, Shaonti Goswami, Md. Azizul Hakim

Automatic Speech Recognition (ASR) has undergone a profound transformation over the past decade, driven by advances in deep learning. This survey provides a comprehensive overview of the modern era of ASR, charting its evolution from traditional hybrid systems, such as Gaussian Mixture Model-Hidden Markov Models (GMM-HMMs) and Deep Neural Network-HMMs (DNN-HMMs), to the now-dominant end-to-end neural architectures. We systematically review the foundational end-to-end paradigms: Connectionist Temporal Classification (CTC), attention-based encoder-decoder models, and the Recurrent Neural Network Transducer (RNN-T), which established the groundwork for fully integrated speech-to-text systems. We then detail the subsequent architectural shift towards Transformer and Conformer models, which leverage self-attention to capture long-range dependencies with high computational efficiency. A central theme of this survey is the parallel revolution in training paradigms. We examine the progression from fully supervised learning, augmented by techniques like SpecAugment, to the rise of self-supervised learning (SSL) with foundation models such as wav2vec 2.0, which drastically reduce the reliance on transcribed data. Furthermore, we analyze the impact of largescale, weakly supervised models like Whisper, which achieve unprecedented robustness through massive data diversity. The paper also covers essential ecosystem components, including key datasets and benchmarks (e.g., LibriSpeech, Switchboard, CHiME), standard evaluation metrics (e.g., Word Error Rate), and critical considerations for real-world deployment, such as streaming inference, on-device efficiency, and the ethical imperatives of fairness and robustness. We conclude by outlining open challenges and future research directions.

♻ ☆ $\texttt{AVROBUSTBENCH}$: Benchmarking the Robustness of Audio-Visual Recognition Models at Test-Time NeurIPS 2025

Sarthak Kumar Maharana, Saksham Singh Kushwaha, Baoming Zhang, Adrian Rodriguez, Songtao Wei, Yapeng Tian, Yunhui Guo

While recent audio-visual models have demonstrated impressive performance, their robustness to distributional shifts at test-time remains not fully understood. Existing robustness benchmarks mainly focus on single modalities, making them insufficient for thoroughly assessing the robustness of audio-visual models. Motivated by real-world scenarios where shifts can occur $\textit{simultaneously}$ in both audio and visual modalities, we introduce $\texttt{AVROBUSTBENCH}$, a comprehensive benchmark designed to evaluate the test-time robustness of audio-visual recognition models. $\texttt{AVROBUSTBENCH}$ comprises four audio-visual benchmark datasets, $\texttt{AUDIOSET-2C}$, $\texttt{VGGSOUND-2C}$, $\texttt{KINETICS-2C}$, and $\texttt{EPICKITCHENS-2C}$, each incorporating 75 bimodal audio-visual corruptions that are $\textit{co-occurring}$ and $\textit{correlated}$. Through extensive evaluations, we observe that state-of-the-art supervised and self-supervised audio-visual models exhibit declining robustness as corruption severity increases. Furthermore, online test-time adaptation (TTA) methods, on $\texttt{VGGSOUND-2C}$ and $\texttt{KINETICS-2C}$, offer minimal improvements in performance under bimodal corruptions. We further propose $\texttt{AV2C}$, a simple TTA approach enabling on-the-fly cross-modal fusion by penalizing high-entropy samples, which achieves improvements on $\texttt{VGGSOUND-2C}$. We hope that $\texttt{AVROBUSTBENCH}$ will steer the development of more effective and robust audio-visual TTA approaches. Our code is available $\href{https://github.com/sarthaxxxxx/AV-C-Robustness-Benchmark}{here}$.

comment: 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Track on Datasets and Benchmarks

♻ ☆ Enhancing Speaker Verification with w2v-BERT 2.0 and Knowledge Distillation guided Structured Pruning

Ze Li, Ming Cheng, Ming Li

Large-scale self-supervised Pre-Trained Models (PTMs) have shown significant improvements in the speaker verification (SV) task by providing rich feature representations. In this paper, we utilize w2v-BERT 2.0, a model with approximately 600 million parameters trained on 4.5 million hours of unlabeled data across 143 languages, for the SV task. The MFA structure with Layer Adapter is employed to process the multi-layer feature outputs from the PTM and extract speaker embeddings. Additionally, we incorporate LoRA for efficient fine-tuning. Our model achieves state-of-the-art results with 0.12% and 0.55% EER on the Vox1-O and Vox1-H test sets, respectively. Furthermore, we apply knowledge distillation guided structured pruning, reducing the model size by 80% while achieving only a 0.04% EER degradation. Source code and models are released at https://github.com/ZXHY-82/w2v-BERT-2.0_SV.

♻ ☆ Audio Does Matter: Importance-Aware Multi-Granularity Fusion for Video Moment Retrieval ACM MM 2025

Junan Lin, Daizong Liu, Xianke Chen, Xiaoye Qu, Xun Yang, Jixiang Zhu, Sanyuan Zhang, Jianfeng Dong

Video Moment Retrieval (VMR) aims to retrieve a specific moment semantically related to the given query. To tackle this task, most existing VMR methods solely focus on the visual and textual modalities while neglecting the complementary but important audio modality. Although a few recent works try to tackle the joint audio-vision-text reasoning, they treat all modalities equally and simply embed them without fine-grained interaction for moment retrieval. These designs are counter-practical as: Not all audios are helpful for video moment retrieval, and the audio of some videos may be complete noise or background sound that is meaningless to the moment determination. To this end, we propose a novel Importance-aware Multi-Granularity fusion model (IMG), which learns to dynamically and selectively aggregate the audio-vision-text contexts for VMR. Specifically, after integrating the textual guidance with vision and audio separately, we first design a pseudo-label-supervised audio importance predictor that predicts the importance score of the audio, and accordingly assigns weights to mitigate the interference caused by noisy audio. Then, we design a multi-granularity audio fusion module that adaptively fuses audio and visual modalities at local-, event-, and global-level, fully capturing their complementary contexts. We further propose a cross-modal knowledge distillation strategy to address the challenge of missing audio modality during inference. To evaluate our method, we further construct a new VMR dataset, i.e., Charades-AudioMatter, where audio-related samples are manually selected and re-organized from the original Charades-STA to validate the model's capability in utilizing audio modality. Extensive experiments validate the effectiveness of our method, achieving state-of-the-art with audio-video fusion in VMR methods. Our code is available at https://github.com/HuiGuanLab/IMG.

comment: Accepted to ACM MM 2025

♻ ☆ Benchmarking and Bridging Emotion Conflicts for Multimodal Emotion Reasoning

Zhiyuan Han, Beier Zhu, Yanlong Xu, Peipei Song, Xun Yang

Despite their strong performance in multimodal emotion reasoning, existing Multimodal Large Language Models (MLLMs) often overlook the scenarios involving emotion conflicts, where emotional cues from different modalities are inconsistent. To fill this gap, we first introduce CA-MER, a new benchmark designed to examine MLLMs under realistic emotion conflicts. It consists of three subsets: video-aligned, audio-aligned, and consistent, where only one or all modalities reflect the true emotion. However, evaluations on our CA-MER reveal that current state-of-the-art emotion MLLMs systematically over-rely on audio signal during emotion conflicts, neglecting critical cues from visual modality. To mitigate this bias, we propose MoSEAR, a parameter-efficient framework that promotes balanced modality integration. MoSEAR consists of two modules: (1)MoSE, modality-specific experts with a regularized gating mechanism that reduces modality bias in the fine-tuning heads; and (2)AR, an attention reallocation mechanism that rebalances modality contributions in frozen backbones during inference. Our framework offers two key advantages: it mitigates emotion conflicts and improves performance on consistent samples-without incurring a trade-off between audio and visual modalities. Experiments on multiple benchmarks-including MER2023, EMER, DFEW, and our CA-MER-demonstrate that MoSEAR achieves state-of-the-art performance, particularly under modality conflict conditions.

comment: ACM Multimedia 2025 Oral Code: https://github.com/ZhiyuanHan-Aaron/MoSEAR Project Page: https://zhiyuanhan-aaron.github.io/MoSEAR-page/

♻ ☆ PicoAudio2: Temporal Controllable Text-to-Audio Generation with Natural Language Description

Zihao Zheng, Zeyu Xie, Xuenan Xu, Wen Wu, Chao Zhang, Mengyue Wu

While recent work in controllable text-to-audio (TTA) generation has achieved fine-grained control through timestamp conditioning, its scope remains limited by audio quality and input format. These models often suffer from poor audio quality in real datasets due to sole reliance on synthetic data. Moreover, some models are constrained to a closed vocabulary of sound events, preventing them from controlling audio generation for open-ended, free-text queries. This paper introduces PicoAudio2, a framework that advances temporal-controllable TTA by mitigating these data and architectural limitations. Specifically, we use a grounding model to annotate event timestamps of real audio-text datasets to curate temporally-strong real data, in addition to simulation data from existing works. The model is trained on the combination of real and simulation data. Moreover, we propose an enhanced architecture that integrates the fine-grained information from a timestamp matrix with coarse-grained free-text input. Experiments show that PicoAudio2 exhibits superior performance in terms of temporal controllability and audio quality.

comment: Demo page: https://HiRookie9.github.io/PicoAudio2-Page

☆ ProGress: Structured Music Generation via Graph Diffusion and Hierarchical Music Analysis

Stephen Ni-Hahn, Chao Péter Yang, Mingchen Ma, Cynthia Rudin, Simon Mak, Yue Jiang

Artificial Intelligence (AI) for music generation is undergoing rapid developments, with recent symbolic models leveraging sophisticated deep learning and diffusion model algorithms. One drawback with existing models is that they lack structural cohesion, particularly on harmonic-melodic structure. Furthermore, such existing models are largely "black-box" in nature and are not musically interpretable. This paper addresses these limitations via a novel generative music framework that incorporates concepts of Schenkerian analysis (SchA) in concert with a diffusion modeling framework. This framework, which we call ProGress (Prolongation-enhanced DiGress), adapts state-of-the-art deep models for discrete diffusion (in particular, the DiGress model of Vignac et al., 2023) for interpretable and structured music generation. Concretely, our contributions include 1) novel adaptations of the DiGress model for music generation, 2) a novel SchA-inspired phrase fusion methodology, and 3) a framework allowing users to control various aspects of the generation process to create coherent musical compositions. Results from human experiments suggest superior performance to existing state-of-the-art methods.

☆ Peransformer: Improving Low-informed Expressive Performance Rendering with Score-aware Discriminator SC 2025

Xian He, Wei Zeng, Ye Wang

Highly-informed Expressive Performance Rendering (EPR) systems transform music scores with rich musical annotations into human-like expressive performance MIDI files. While these systems have achieved promising results, the availability of detailed music scores is limited compared to MIDI files and are less flexible to work with using a digital audio workstation (DAW). Recent advancements in low-informed EPR systems offer a more accessible alternative by directly utilizing score-derived MIDI as input, but these systems often exhibit suboptimal performance. Meanwhile, existing works are evaluated with diverse automatic metrics and data formats, hindering direct objective comparisons between EPR systems. In this study, we introduce Peransformer, a transformer-based low-informed EPR system designed to bridge the gap between low-informed and highly-informed EPR systems. Our approach incorporates a score-aware discriminator that leverages the underlying score-derived MIDI files and is trained on a score-to-performance paired, note-to-note aligned MIDI dataset. Experimental results demonstrate that Peransformer achieves state-of-the-art performance among low-informed systems, as validated by subjective evaluations. Furthermore, we extend existing automatic evaluation metrics for EPR systems and introduce generalized EPR metrics (GEM), enabling more direct, accurate, and reliable comparisons across EPR systems.

comment: 6 pages, 3 figures, accepted by APSIPA ASC 2025

☆ Chord Colourizer: A Near Real-Time System for Visualizing Musical Key

Paul Haimes

This paper introduces Chord Colourizer, a near real-time system that detects the musical key of an audio signal and visually represents it through a novel graphical user interface (GUI). The system assigns colours to musical notes based on Isaac Newton's original colour wheel, preserving historical links between pitch and hue, and also integrates an Arduino-controlled LED display using 3D-printed star-shaped diffusers to offer a physical ambient media representation. The method employs Constant-Q Transform (CQT) chroma features for chord estimation and visualization, followed by threshold-based filtering and tonal enhancement to isolate the root, third, and fifth. A confidence score is computed for each detection to ensure reliability, and only chords with moderate to very strong certainty are visualized. The graphical interface dynamically updates a colour-coded keyboard layout, while the LED display provides the same colour information via spatial feedback. This multi-modal system enhances user interaction with harmonic content, offering innovative possibilities for education and artistic performance. Limitations include slight latency and the inability to detect extended chords, which future development will aim to address through refined filtering, adaptive thresholds, and support for more complex harmonies such as sevenths and augmented chords. Future work will also explore integration with alternative visualization styles, and the comparison of audio analysis libraries to improve detection speed and precision. Plans also include formal user testing to evaluate perception, usability, and cross-cultural interpretations of colour-pitch mappings.

comment: Author copy. This paper is in press for presentation at ADADA 2025. Please cite as: Haimes, P. (in press). Chord Colourizer: A near real-time system for visualizing musical key. In Proceedings of the 23rd International Conference of Asia Digital Art and Design Association (ADADA)

☆ MTP-S2UT: Enhancing Speech-to-Speech Translation Quality with Multi-token Prediction

Jianjin Wang, Runsong Zhao, Xiaoqian Liu, Yuan Ge, Ziqiang Xu, Tong Xiao, Shengxiang Gao, Zhengtao Yu, Jingbo Zhu

Current direct speech-to-speech translation methods predominantly employ speech tokens as intermediate representations. However, a single speech token is not dense in semantics, so we generally need multiple tokens to express a complete semantic unit. To address this limitation, we introduce multi-token prediction (MTP) loss into speech-to-unit translation (S2UT) models, enabling models to predict multiple subsequent tokens at each position, thereby capturing more complete semantics and enhancing information density per position. Initial MTP implementations apply the loss at the final layer, which improves output representation but initiates information enrichment too late. We hypothesize that advancing the information enrichment process to intermediate layers can achieve earlier and more effective enhancement of hidden representation. Consequently, we propose MTP-S2UT loss, applying MTP loss to hidden representation where CTC loss is computed. Experiments demonstrate that all MTP loss variants consistently improve the quality of S2UT translation, with MTP-S2UT achieving the best performance.

comment: Copyright 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works

☆ Automatic Speech Recognition in the Modern Era: Architectures, Training, and Evaluation

Md. Nayeem, Md Shamse Tabrej, Kabbojit Jit Deb, Shaonti Goswami, Md. Azizul Hakim

Automatic Speech Recognition (ASR) has undergone a profound transformation over the past decade, driven by advances in deep learning. This survey provides a comprehensive overview of the modern era of ASR, charting its evolution from traditional hybrid systems, such as Gaussian Mixture Model-Hidden Markov Models (GMM-HMMs) and Deep Neural Network-HMMs (DNN-HMMs), to the now-dominant end-to-end neural architectures. We systematically review the foundational end-to-end paradigms: Connectionist Temporal Classification (CTC), attention-based encoder-decoder models, and the Recurrent Neural Network Transducer (RNN-T), which established the groundwork for fully integrated speech-to-text systems. We then detail the subsequent architectural shift towards Transformer and Conformer models, which leverage self-attention to capture long-range dependencies with high computational efficiency. A central theme of this survey is the parallel revolution in training paradigms. We examine the progression from fully supervised learning, augmented by techniques like SpecAugment, to the rise of self-supervised learning (SSL) with foundation models such as wav2vec 2.0, which drastically reduce the reliance on transcribed data. Furthermore, we analyze the impact of largescale, weakly supervised models like Whisper, which achieve unprecedented robustness through massive data diversity. The paper also covers essential ecosystem components, including key datasets and benchmarks (e.g., LibriSpeech, Switchboard, CHiME), standard evaluation metrics (e.g., Word Error Rate), and critical considerations for real-world deployment, such as streaming inference, on-device efficiency, and the ethical imperatives of fairness and robustness. We conclude by outlining open challenges and future research directions.

♻ ☆ $\texttt{AVROBUSTBENCH}$: Benchmarking the Robustness of Audio-Visual Recognition Models at Test-Time NeurIPS 2025

Sarthak Kumar Maharana, Saksham Singh Kushwaha, Baoming Zhang, Adrian Rodriguez, Songtao Wei, Yapeng Tian, Yunhui Guo

While recent audio-visual models have demonstrated impressive performance, their robustness to distributional shifts at test-time remains not fully understood. Existing robustness benchmarks mainly focus on single modalities, making them insufficient for thoroughly assessing the robustness of audio-visual models. Motivated by real-world scenarios where shifts can occur $\textit{simultaneously}$ in both audio and visual modalities, we introduce $\texttt{AVROBUSTBENCH}$, a comprehensive benchmark designed to evaluate the test-time robustness of audio-visual recognition models. $\texttt{AVROBUSTBENCH}$ comprises four audio-visual benchmark datasets, $\texttt{AUDIOSET-2C}$, $\texttt{VGGSOUND-2C}$, $\texttt{KINETICS-2C}$, and $\texttt{EPICKITCHENS-2C}$, each incorporating 75 bimodal audio-visual corruptions that are $\textit{co-occurring}$ and $\textit{correlated}$. Through extensive evaluations, we observe that state-of-the-art supervised and self-supervised audio-visual models exhibit declining robustness as corruption severity increases. Furthermore, online test-time adaptation (TTA) methods, on $\texttt{VGGSOUND-2C}$ and $\texttt{KINETICS-2C}$, offer minimal improvements in performance under bimodal corruptions. We further propose $\texttt{AV2C}$, a simple TTA approach enabling on-the-fly cross-modal fusion by penalizing high-entropy samples, which achieves improvements on $\texttt{VGGSOUND-2C}$. We hope that $\texttt{AVROBUSTBENCH}$ will steer the development of more effective and robust audio-visual TTA approaches. Our code is available $\href{https://github.com/sarthaxxxxx/AV-C-Robustness-Benchmark}{here}$.

comment: 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Track on Datasets and Benchmarks

♻ ☆ Enhancing Speaker Verification with w2v-BERT 2.0 and Knowledge Distillation guided Structured Pruning

Ze Li, Ming Cheng, Ming Li

Large-scale self-supervised Pre-Trained Models (PTMs) have shown significant improvements in the speaker verification (SV) task by providing rich feature representations. In this paper, we utilize w2v-BERT 2.0, a model with approximately 600 million parameters trained on 4.5 million hours of unlabeled data across 143 languages, for the SV task. The MFA structure with Layer Adapter is employed to process the multi-layer feature outputs from the PTM and extract speaker embeddings. Additionally, we incorporate LoRA for efficient fine-tuning. Our model achieves state-of-the-art results with 0.12% and 0.55% EER on the Vox1-O and Vox1-H test sets, respectively. Furthermore, we apply knowledge distillation guided structured pruning, reducing the model size by 80% while achieving only a 0.04% EER degradation. Source code and models are released at https://github.com/ZXHY-82/w2v-BERT-2.0_SV.

♻ ☆ Audio Does Matter: Importance-Aware Multi-Granularity Fusion for Video Moment Retrieval ACM MM 2025

Junan Lin, Daizong Liu, Xianke Chen, Xiaoye Qu, Xun Yang, Jixiang Zhu, Sanyuan Zhang, Jianfeng Dong

Video Moment Retrieval (VMR) aims to retrieve a specific moment semantically related to the given query. To tackle this task, most existing VMR methods solely focus on the visual and textual modalities while neglecting the complementary but important audio modality. Although a few recent works try to tackle the joint audio-vision-text reasoning, they treat all modalities equally and simply embed them without fine-grained interaction for moment retrieval. These designs are counter-practical as: Not all audios are helpful for video moment retrieval, and the audio of some videos may be complete noise or background sound that is meaningless to the moment determination. To this end, we propose a novel Importance-aware Multi-Granularity fusion model (IMG), which learns to dynamically and selectively aggregate the audio-vision-text contexts for VMR. Specifically, after integrating the textual guidance with vision and audio separately, we first design a pseudo-label-supervised audio importance predictor that predicts the importance score of the audio, and accordingly assigns weights to mitigate the interference caused by noisy audio. Then, we design a multi-granularity audio fusion module that adaptively fuses audio and visual modalities at local-, event-, and global-level, fully capturing their complementary contexts. We further propose a cross-modal knowledge distillation strategy to address the challenge of missing audio modality during inference. To evaluate our method, we further construct a new VMR dataset, i.e., Charades-AudioMatter, where audio-related samples are manually selected and re-organized from the original Charades-STA to validate the model's capability in utilizing audio modality. Extensive experiments validate the effectiveness of our method, achieving state-of-the-art with audio-video fusion in VMR methods. Our code is available at https://github.com/HuiGuanLab/IMG.

comment: Accepted to ACM MM 2025

♻ ☆ SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision

Chunbo Hao, Ruibin Yuan, Jixun Yao, Qixin Deng, Xinyi Bai, Wei Xue, Lei Xie

Music structure analysis (MSA) underpins music understanding and controllable generation, yet progress has been limited by small, inconsistent corpora. We present SongFormer, a scalable framework that learns from heterogeneous supervision. SongFormer (i) fuses short- and long-window self-supervised audio representations to capture both fine-grained and long-range dependencies, and (ii) introduces a learned source embedding to enable training with partial, noisy, and schema-mismatched labels. To support scaling and fair evaluation, we release SongFormDB, the largest MSA corpus to date (over 10k tracks spanning languages and genres), and SongFormBench, a 300-song expert-verified benchmark. On SongFormBench, SongFormer sets a new state of the art in strict boundary detection (HR.5F) and achieves the highest functional label accuracy, while remaining computationally efficient; it surpasses strong baselines and Gemini 2.5 Pro on these metrics and remains competitive under relaxed tolerance (HR3F). Code, datasets, and model are publicly available.

♻ ☆ Benchmarking and Bridging Emotion Conflicts for Multimodal Emotion Reasoning

Zhiyuan Han, Beier Zhu, Yanlong Xu, Peipei Song, Xun Yang

Despite their strong performance in multimodal emotion reasoning, existing Multimodal Large Language Models (MLLMs) often overlook the scenarios involving emotion conflicts, where emotional cues from different modalities are inconsistent. To fill this gap, we first introduce CA-MER, a new benchmark designed to examine MLLMs under realistic emotion conflicts. It consists of three subsets: video-aligned, audio-aligned, and consistent, where only one or all modalities reflect the true emotion. However, evaluations on our CA-MER reveal that current state-of-the-art emotion MLLMs systematically over-rely on audio signal during emotion conflicts, neglecting critical cues from visual modality. To mitigate this bias, we propose MoSEAR, a parameter-efficient framework that promotes balanced modality integration. MoSEAR consists of two modules: (1)MoSE, modality-specific experts with a regularized gating mechanism that reduces modality bias in the fine-tuning heads; and (2)AR, an attention reallocation mechanism that rebalances modality contributions in frozen backbones during inference. Our framework offers two key advantages: it mitigates emotion conflicts and improves performance on consistent samples-without incurring a trade-off between audio and visual modalities. Experiments on multiple benchmarks-including MER2023, EMER, DFEW, and our CA-MER-demonstrate that MoSEAR achieves state-of-the-art performance, particularly under modality conflict conditions.

comment: ACM Multimedia 2025 Oral Code: https://github.com/ZhiyuanHan-Aaron/MoSEAR Project Page: https://zhiyuanhan-aaron.github.io/MoSEAR-page/

♻ ☆ PicoAudio2: Temporal Controllable Text-to-Audio Generation with Natural Language Description

Zihao Zheng, Zeyu Xie, Xuenan Xu, Wen Wu, Chao Zhang, Mengyue Wu

While recent work in controllable text-to-audio (TTA) generation has achieved fine-grained control through timestamp conditioning, its scope remains limited by audio quality and input format. These models often suffer from poor audio quality in real datasets due to sole reliance on synthetic data. Moreover, some models are constrained to a closed vocabulary of sound events, preventing them from controlling audio generation for open-ended, free-text queries. This paper introduces PicoAudio2, a framework that advances temporal-controllable TTA by mitigating these data and architectural limitations. Specifically, we use a grounding model to annotate event timestamps of real audio-text datasets to curate temporally-strong real data, in addition to simulation data from existing works. The model is trained on the combination of real and simulation data. Moreover, we propose an enhanced architecture that integrates the fine-grained information from a timestamp matrix with coarse-grained free-text input. Experiments show that PicoAudio2 exhibits superior performance in terms of temporal controllability and audio quality.

comment: Demo page: https://HiRookie9.github.io/PicoAudio2-Page

☆ Phase-Aware Deep Learning with Complex-Valued CNNs for Audio Signal Applications

Naman Agrawal

This study explores the design and application of Complex-Valued Convolutional Neural Networks (CVCNNs) in audio signal processing, with a focus on preserving and utilizing phase information often neglected in real-valued networks. We begin by presenting the foundational theoretical concepts of CVCNNs, including complex convolutions, pooling layers, Wirtinger-based differentiation, and various complex-valued activation functions. These are complemented by critical adaptations of training techniques, including complex batch normalization and weight initialization schemes, to ensure stability in training dynamics. Empirical evaluations are conducted across three stages. First, CVCNNs are benchmarked on standard image datasets, where they demonstrate competitive performance with real-valued CNNs, even under synthetic complex perturbations. Although our focus is audio signal processing, we first evaluate CVCNNs on image datasets to establish baseline performance and validate training stability before applying them to audio tasks. In the second experiment, we focus on audio classification using Mel-Frequency Cepstral Coefficients (MFCCs). CVCNNs trained on real-valued MFCCs slightly outperform real CNNs, while preserving phase in input workflows highlights challenges in exploiting phase without architectural modifications. Finally, a third experiment introduces GNNs to model phase information via edge weighting, where the inclusion of phase yields measurable gains in both binary and multi-class genre classification. These results underscore the expressive capacity of complex-valued architectures and confirm phase as a meaningful and exploitable feature in audio processing applications. While current methods show promise, especially with activations like cardioid, future advances in phase-aware design will be essential to leverage the potential of complex representations in neural networks.

☆ Accent-Invariant Automatic Speech Recognition via Saliency-Driven Spectrogram Masking ICASSP 2026

Mohammad Hossein Sameti, Sepehr Harfi Moridani, Ali Zarean, Hossein Sameti

Pre-trained transformer-based models have significantly advanced automatic speech recognition (ASR), yet they remain sensitive to accent and dialectal variations, resulting in elevated word error rates (WER) in linguistically diverse languages such as English and Persian. To address this challenge, we propose an accent-invariant ASR framework that integrates accent and dialect classification into the recognition pipeline. Our approach involves training a spectrogram-based classifier to capture accent-specific cues, masking the regions most influential to its predictions, and using the masked spectrograms for data augmentation. This enhances the robustness of ASR models against accent variability. We evaluate the method using both English and Persian speech. For Persian, we introduce a newly collected dataset spanning multiple regional accents, establishing the first systematic benchmark for accent variation in Persian ASR that fills a critical gap in multilingual speech research and provides a foundation for future studies on low-resource, linguistically diverse languages. Experimental results with the Whisper model demonstrate that our masking and augmentation strategy yields substantial WER reductions in both English and Persian settings, confirming the effectiveness of the approach. This research advances the development of multilingual ASR systems that are resilient to accent and dialect diversity. Code and dataset are publicly available at: https://github.com/MH-Sameti/Accent_invariant_ASR

comment: Submitted to ICASSP 2026

☆ WildElder: A Chinese Elderly Speech Dataset from the Wild with Fine-Grained Manual Annotations

Hui Wang, Jiaming Zhou, Jiabei He, Haoqin Sun, Yong Qin

Elderly speech poses unique challenges for automatic processing due to age-related changes such as slower articulation and vocal tremors. Existing Chinese datasets are mostly recorded in controlled environments, limiting their diversity and real-world applicability. To address this gap, we present WildElder, a Mandarin elderly speech corpus collected from online videos and enriched with fine-grained manual annotations, including transcription, speaker age, gender, and accent strength. Combining the realism of in-the-wild data with expert curation, WildElder enables robust research on automatic speech recognition and speaker profiling. Experimental results reveal both the difficulties of elderly speech recognition and the potential of WildElder as a challenging new benchmark. The dataset and code are available at https://github.com/NKU-HLT/WildElder.

☆ SynthVC: Leveraging Synthetic Data for End-to-End Low Latency Streaming Voice Conversion SC2025

Zhao Guo, Ziqian Ning, Guobin Ma, Lei Xie

Voice Conversion (VC) aims to modify a speaker's timbre while preserving linguistic content. While recent VC models achieve strong performance, most struggle in real-time streaming scenarios due to high latency, dependence on ASR modules, or complex speaker disentanglement, which often results in timbre leakage or degraded naturalness. We present SynthVC, a streaming end-to-end VC framework that directly learns speaker timbre transformation from synthetic parallel data generated by a pre-trained zero-shot VC model. This design eliminates the need for explicit content-speaker separation or recognition modules. Built upon a neural audio codec architecture, SynthVC supports low-latency streaming inference with high output fidelity. Experimental results show that SynthVC outperforms baseline streaming VC systems in both naturalness and speaker similarity, achieving an end-to-end latency of just 77.1 ms.

comment: Accepted by NCMMSC2025

☆ Effects of automotive microphone frequency response characteristics and noise conditions on speech and ASR quality -- an experimental evaluation

Michele Buccoli, Yu Du, Jacob Soendergaard, Simone Shawn Cazzaniga

Upon choosing microphones for automotive hands-free communication or Automatic Speech Recognition (ASR) applications, OEMs typically specify wideband, super wideband or even fullband requirements following established standard recommendations (e.g., ITU-P.1110, ITU-P.1120). In practice, it is often challenging to achieve the preferred bandwidth for an automotive microphone when considering limitations and constraints on microphone placement inside the cabin, and the automotive grade environmental robustness requirements. On the other hand, there seems to be no consensus or sufficient data on the effect of each microphone characteristic on the actual performance. As an attempt to answer this question, we used noise signals recorded in real vehicles and under various driving conditions to experimentally study the relationship between the microphones' characteristics and the final audio quality of speech communication and performance of ASR engines. We focus on how variations in microphone bandwidth and amplitude frequency response shapes affect the perceptual speech quality. The speech quality results are compared by using ETSI TS 103 281 metrics (S-MOS, N-MOS, G-MOS) and ancillary metrics such as SNR. The ASR results are evaluated with standard metrics such as Word Error Rate (WER). Findings from this study provide knowledge in the understanding of what microphone frequency response characteristics are more relevant for audio quality and choice of proper microphone specifications, particularly for automotive applications.

☆ Unsupervised lexicon learning from speech is limited by representations rather than clustering ICASSP 2026

Danel Adendorff, Simon Malan, Herman Kamper

Zero-resource word segmentation and clustering systems aim to tokenise speech into word-like units without access to text labels. Despite progress, the induced lexicons are still far from perfect. In an idealised setting with gold word boundaries, we ask whether performance is limited by the representation of word segments, or by the clustering methods that group them into word-like types. We combine a range of self-supervised speech features (continuous/discrete, frame/word-level) with different clustering methods (K-means, hierarchical, graph-based) on English and Mandarin data. The best system uses graph clustering with dynamic time warping on continuous features. Faster alternatives use graph clustering with cosine distance on averaged continuous features or edit distance on discrete unit sequences. Through controlled experiments that isolate either the representations or the clustering method, we demonstrate that representation variability across segments of the same word type -- rather than clustering -- is the primary factor limiting performance.

comment: Submitted to ICASSP 2026

☆ FLToP CTC: Frame-Level Token Pruning via Relative Threshold for Efficient and Memory-Saving Decoding on Diverse Platforms

Atul Shree, Harshith Jupuru

CTC-based ASR systems face computational and memory bottlenecks in resource-limited environments. Traditional CTC decoders, requiring up to 90% of processing time in systems (e.g., wav2vec2-large on L4 GPUs), face inefficiencies due to exhaustive token-level operations. This paper introduces Frame Level Token Pruning for Connectionist Temporal Classification (FLToP CTC), a novel decoding algorithm that employs frame-level token pruning guided by a relative threshold probability. By dynamically eliminating low-probability tokens per frame, FLToP CTC reduces compute and memory demands while maintaining negligible WER degradation. On LibriSpeech, FLToP CTC achieves a 10.5x runtime speedup and 2.78x memory reduction versus standard CTC decoders. Its simplicity enables seamless integration into CTC decoders across platforms (CPUs, GPUs, etc.). FLToP CTC addresses CTC bottlenecks, offering scalability for resource-limited environments and realtime applications, enhancing speech recognition accessibility and efficiency.

comment: 5 pages, 5 figures

☆ Emotion-Disentangled Embedding Alignment for Noise-Robust and Cross-Corpus Speech Emotion Recognition

Upasana Tiwari, Rupayan Chakraborty, Sunil Kumar Kopparapu

Effectiveness of speech emotion recognition in real-world scenarios is often hindered by noisy environments and variability across datasets. This paper introduces a two-step approach to enhance the robustness and generalization of speech emotion recognition models through improved representation learning. First, our model employs EDRL (Emotion-Disentangled Representation Learning) to extract class-specific discriminative features while preserving shared similarities across emotion categories. Next, MEA (Multiblock Embedding Alignment) refines these representations by projecting them into a joint discriminative latent subspace that maximizes covariance with the original speech input. The learned EDRL-MEA embeddings are subsequently used to train an emotion classifier using clean samples from publicly available datasets, and are evaluated on unseen noisy and cross-corpus speech samples. Improved performance under these challenging conditions demonstrates the effectiveness of the proposed method.

comment: 13 pages, 1 figure

☆ MMAudioSep: Taming Video-to-Audio Generative Model Towards Video/Text-Queried Sound Separation

Akira Takahashi, Shusuke Takahashi, Yuki Mitsufuji

We introduce MMAudioSep, a generative model for video/text-queried sound separation that is founded on a pretrained video-to-audio model. By leveraging knowledge about the relationship between video/text and audio learned through a pretrained audio generative model, we can train the model more efficiently, i.e., the model does not need to be trained from scratch. We evaluate the performance of MMAudioSep by comparing it to existing separation models, including models based on both deterministic and generative approaches, and find it is superior to the baseline models. Furthermore, we demonstrate that even after acquiring functionality for sound separation via fine-tuning, the model retains the ability for original video-to-audio generation. This highlights the potential of foundational sound generation models to be adopted for sound-related downstream tasks. Our code is available at https://github.com/sony/mmaudiosep.

comment: 4 pages, 4 figures, 2 tables

☆ O_O-VC: Synthetic Data-Driven One-to-One Alignment for Any-to-Any Voice Conversion EMNLP 2025

Huu Tuong Tu, Huan Vu, cuong tien nguyen, Dien Hy Ngo, Nguyen Thi Thu Trang

Traditional voice conversion (VC) methods typically attempt to separate speaker identity and linguistic information into distinct representations, which are then combined to reconstruct the audio. However, effectively disentangling these factors remains challenging, often leading to information loss during training. In this paper, we propose a new approach that leverages synthetic speech data generated by a high-quality, pretrained multispeaker text-to-speech (TTS) model. Specifically, synthetic data pairs that share the same linguistic content but differ in speaker identity are used as input-output pairs to train the voice conversion model. This enables the model to learn a direct mapping between source and target voices, effectively capturing speaker-specific characteristics while preserving linguistic content. Additionally, we introduce a flexible training strategy for any-to-any voice conversion that generalizes well to unseen speakers and new languages, enhancing adaptability and performance in zero-shot scenarios. Our experiments show that our proposed method achieves a 16.35% relative reduction in word error rate and a 5.91% improvement in speaker cosine similarity, outperforming several state-of-the-art methods. Voice conversion samples can be accessed at: https://oovc-emnlp-2025.github.io/

comment: EMNLP 2025

☆ Déréverbération non-supervisée de la parole par modèle hybride

Louis Bahrman, Mathieu Fontaine, Gaël Richard

This paper introduces a new training strategy to improve speech dereverberation systems in an unsupervised manner using only reverberant speech. Most existing algorithms rely on paired dry/reverberant data, which is difficult to obtain. Our approach uses limited acoustic information, like the reverberation time (RT60), to train a dereverberation system. Experimental results demonstrate that our method achieves more consistent performance across various objective metrics than the state-of-the-art.

comment: in French language

☆ DiTSinger: Scaling Singing Voice Synthesis with Diffusion Transformer and Implicit Alignment

Zongcai Du, Guilin Deng, Xiaofeng Guo, Xin Gao, Linke Li, Kaichang Cheng, Fubo Han, Siyu Yang, Peng Liu, Pan Zhong, Qiang Fu

Recent progress in diffusion-based Singing Voice Synthesis (SVS) demonstrates strong expressiveness but remains limited by data scarcity and model scalability. We introduce a two-stage pipeline: a compact seed set of human-sung recordings is constructed by pairing fixed melodies with diverse LLM-generated lyrics, and melody-specific models are trained to synthesize over 500 hours of high-quality Chinese singing data. Building on this corpus, we propose DiTSinger, a Diffusion Transformer with RoPE and qk-norm, systematically scaled in depth, width, and resolution for enhanced fidelity. Furthermore, we design an implicit alignment mechanism that obviates phoneme-level duration labels by constraining phoneme-to-acoustic attention within character-level spans, thereby improving robustness under noisy or uncertain alignments. Extensive experiments validate that our approach enables scalable, alignment-free, and high-fidelity SVS.

comment: under review

☆ VM-UNSSOR: Unsupervised Neural Speech Separation Enhanced by Higher-SNR Virtual Microphone Arrays

Shulin He, Zhong-Qiu Wang

Blind speech separation (BSS) aims to recover multiple speech sources from multi-channel, multi-speaker mixtures under unknown array geometry and room impulse responses. In unsupervised setup where clean target speech is not available for model training, UNSSOR proposes a mixture consistency (MC) loss for training deep neural networks (DNN) on over-determined training mixtures to realize unsupervised speech separation. However, when the number of microphones of the training mixtures decreases, the MC constraint weakens and the separation performance falls dramatically. To address this, we propose VM-UNSSOR, augmenting the observed training mixture signals recorded by a limited number of microphones with several higher-SNR virtual-microphone (VM) signals, which are obtained by applying linear spatial demixers (such as IVA and spatial clustering) to the observed training mixtures. As linear projections of the observed mixtures, the virtual-microphone signals can typically increase the SNR of each source and can be leveraged to compute extra MC losses to improve UNSSOR and address the frequency permutation problem in UNSSOR. On the SMS-WSJ dataset, in the over-determined six-microphone, two-speaker separation setup, VM-UNSSOR reaches 17.1 dB SI-SDR, while UNSSOR only obtains 14.7 dB; and in the determined two-microphone, two-speaker case, UNSSOR collapses to -2.7 dB SI-SDR, while VM-UNSSOR achieves 10.7 dB.

☆ ControlAudio: Tackling Text-Guided, Timing-Indicated and Intelligible Audio Generation via Progressive Diffusion Modeling

Yuxuan Jiang, Zehua Chen, Zeqian Ju, Yusheng Dai, Weibei Dou, Jun Zhu

Text-to-audio (TTA) generation with fine-grained control signals, e.g., precise timing control or intelligible speech content, has been explored in recent works. However, constrained by data scarcity, their generation performance at scale is still compromised. In this study, we recast controllable TTA generation as a multi-task learning problem and introduce a progressive diffusion modeling approach, ControlAudio. Our method adeptly fits distributions conditioned on more fine-grained information, including text, timing, and phoneme features, through a step-by-step strategy. First, we propose a data construction method spanning both annotation and simulation, augmenting condition information in the sequence of text, timing, and phoneme. Second, at the model training stage, we pretrain a diffusion transformer (DiT) on large-scale text-audio pairs, achieving scalable TTA generation, and then incrementally integrate the timing and phoneme features with unified semantic representations, expanding controllability. Finally, at the inference stage, we propose progressively guided generation, which sequentially emphasizes more fine-grained information, aligning inherently with the coarse-to-fine sampling nature of DiT. Extensive experiments show that ControlAudio achieves state-of-the-art performance in terms of temporal accuracy and speech clarity, significantly outperforming existing methods on both objective and subjective evaluations. Demo samples are available at: https://control-audio.github.io/Control-Audio.

comment: 18 pages, 8 tables, 5 figures

☆ SeeingSounds: Learning Audio-to-Visual Alignment via Text

Simone Carnemolla, Matteo Pennisi, Chiara Russo, Simone Palazzo, Daniela Giordano, Concetto Spampinato

We introduce SeeingSounds, a lightweight and modular framework for audio-to-image generation that leverages the interplay between audio, language, and vision-without requiring any paired audio-visual data or training on visual generative models. Rather than treating audio as a substitute for text or relying solely on audio-to-text mappings, our method performs dual alignment: audio is projected into a semantic language space via a frozen language encoder, and, contextually grounded into the visual domain using a vision-language model. This approach, inspired by cognitive neuroscience, reflects the natural cross-modal associations observed in human perception. The model operates on frozen diffusion backbones and trains only lightweight adapters, enabling efficient and scalable learning. Moreover, it supports fine-grained and interpretable control through procedural text prompt generation, where audio transformations (e.g., volume or pitch shifts) translate into descriptive prompts (e.g., "a distant thunder") that guide visual outputs. Extensive experiments across standard benchmarks confirm that SeeingSounds outperforms existing methods in both zero-shot and supervised settings, establishing a new state of the art in controllable audio-to-visual generation.

comment: accepted to ACM Multimedia Asia 2025

☆ Serial-Parallel Dual-Path Architecture for Speaking Style Recognition SC2025

Guojian Li, Qijie Shao, Zhixian Zhao, Shuiyuan Wang, Zhonghua Fu, Lei Xie

Speaking Style Recognition (SSR) identifies a speaker's speaking style characteristics from speech. Existing style recognition approaches primarily rely on linguistic information, with limited integration of acoustic information, which restricts recognition accuracy improvements. The fusion of acoustic and linguistic modalities offers significant potential to enhance recognition performance. In this paper, we propose a novel serial-parallel dual-path architecture for SSR that leverages acoustic-linguistic bimodal information. The serial path follows the ASR+STYLE serial paradigm, reflecting a sequential temporal dependency, while the parallel path integrates our designed Acoustic-Linguistic Similarity Module (ALSM) to facilitate cross-modal interaction with temporal simultaneity. Compared to the existing SSR baseline -- the OSUM model, our approach reduces parameter size by 88.4% and achieves a 30.3% improvement in SSR accuracy for eight styles on the test set.

comment: Accepted by NCMMSC2025

☆ Production and Manufacturing of 3D Printed Acoustic Guitars

Timothy Tran, William Schiesser

This research investigates the feasibility of producing affordable, functional acoustic guitars using 3D printing, with a focus on producing structural designs with proper tonal performance. Conducted in collaboration with William Schiesser, the study uses a classical guitar model, chosen for its lower string tension, to evaluate the tonal characteristics of a 3D-printed prototype made from polylactic acid (PLA). Due to the build plate size constraints of the Prusa Mark 4 printer, the guitar body was divided into multiple sections joined with press-fit tolerances and minimal cyanoacrylate adhesive. CAD modeling in Fusion 360 ensured dimensional accuracy in press-fit connections and the overall assembly. Following assembly, the guitar was strung with nylon strings and tested using Audacity software to compare recorded frequencies and notes with standard reference values. Results showed large deviations in lower string frequencies, likely caused by the material choice utilized in printing. Accurate pitches were reached with all strings despite frequency differences through tuning, demonstrating that PLA and modern manufacturing methods can produce affordable, playable acoustic guitars despite inevitable challenges. Further research may investigate alternative plastics for superior frequency matching. This approach holds significant potential for expanding access to quality instruments while reducing reliance on endangered tonewoods, thereby encouraging both sustainable instrument production and increased musical participation. This also creates opportunities for disadvantaged communities where access to musical instruments remains a challenge. Keywords: Luthiery, Stereolithography, 3D-Print, Guitar Making

♻ ☆ Direction Estimation of Sound Sources Using Microphone Arrays and Signal Strength ICSE

Mahdi Ali Pour, Zahra Habibzadeh

Sound-tracking refers to the process of determining the direction from which a sound originates, making it a fundamental component of sound source localization. This capability is essential in a variety of applications, including security systems, acoustic monitoring, and speaker tracking, where accurately identifying the direction of a sound source enables real-time responses, efficient resource allocation, and improved situational awareness. While sound-tracking is closely related to localization, it specifically focuses on identifying the direction of the sound source rather than estimating its exact position in space. Despite its utility, sound-tracking systems face several challenges, such as maintaining directional accuracy and precision, along with the need for sophisticated hardware configurations and complex signal processing algorithms. This paper presents a sound-tracking method using three electret microphones. We estimate the direction of a sound source using a lightweight method that analyzes signals from three strategically placed microphones. By comparing the average power of the received signals, the system infers the most probable direction of the sound. The results indicate that the power level from each microphone effectively determines the sound source direction. Our system employs a straightforward and cost-effective hardware design, ensuring simplicity and affordability in implementation. It achieves a localization error of less than 6 degrees and a precision of 98%. Additionally, its effortless integration with various systems makes it versatile and adaptable. Consequently, this technique presents a robust and reliable solution for sound-tracking and localization, with potential applications spanning diverse domains such as security systems, smart homes, and acoustic monitoring.

comment: Accepted to the 32nd International Conference on Systems Engineering (ICSEng'2025)

♻ ☆ TARO: Timestep-Adaptive Representation Alignment with Onset-Aware Conditioning for Synchronized Video-to-Audio Synthesis ICCV 2025

Tri Ton, Ji Woo Hong, Chang D. Yoo

This paper introduces Timestep-Adaptive Representation Alignment with Onset-Aware Conditioning (TARO), a novel framework for high-fidelity and temporally coherent video-to-audio synthesis. Built upon flow-based transformers, which offer stable training and continuous transformations for enhanced synchronization and audio quality, TARO introduces two key innovations: (1) Timestep-Adaptive Representation Alignment (TRA), which dynamically aligns latent representations by adjusting alignment strength based on the noise schedule, ensuring smooth evolution and improved fidelity, and (2) Onset-Aware Conditioning (OAC), which integrates onset cues that serve as sharp event-driven markers of audio-relevant visual moments to enhance synchronization with dynamic visual events. Extensive experiments on the VGGSound and Landscape datasets demonstrate that TARO outperforms prior methods, achieving relatively 53% lower Frechet Distance (FD), 29% lower Frechet Audio Distance (FAD), and a 97.19% Alignment Accuracy, highlighting its superior audio quality and synchronization precision.

comment: Accepted to ICCV 2025. Please visit our project page at https://triton99.github.io/taro-site/

♻ ☆ From Coarse to Fine: Recursive Audio-Visual Semantic Enhancement for Speech Separation

Ke Xue, Rongfei Fan, Lixin, Dawei Zhao, Chao Zhu, Han Hu

Audio-visual speech separation aims to isolate each speaker's clean voice from mixtures by leveraging visual cues such as lip movements and facial features. While visual information provides complementary semantic guidance, existing methods often underexploit its potential by relying on static visual representations. In this paper, we propose CSFNet, a Coarse-to-Separate-Fine Network that introduces a recursive semantic enhancement paradigm for more effective separation. CSFNet operates in two stages: (1) Coarse Separation, where a first-pass estimation reconstructs a coarse audio waveform from the mixture and visual input; and (2) Fine Separation, where the coarse audio is fed back into an audio-visual speech recognition (AVSR) model together with the visual stream. This recursive process produces more discriminative semantic representations, which are then used to extract refined audio. To further exploit these semantics, we design a speaker-aware perceptual fusion block to encode speaker identity across modalities, and a multi-range spectro-temporal separation network to capture both local and global time-frequency patterns. Extensive experiments on three benchmark datasets and two noisy datasets show that CSFNet achieves state-of-the-art (SOTA) performance, with substantial coarse-to-fine improvements, validating the necessity and effectiveness of our recursive semantic enhancement framework.

♻ ☆ Phonikud: Hebrew Grapheme-to-Phoneme Conversion for Real-Time Text-to-Speech

Yakov Kolani, Maxim Melichov, Cobi Calev, Morris Alper

Real-time text-to-speech (TTS) for Modern Hebrew is challenging due to the language's orthographic complexity. Existing solutions ignore crucial phonetic features such as stress that remain underspecified even when vowel marks are added. To address these limitations, we introduce Phonikud, a lightweight, open-source Hebrew grapheme-to-phoneme (G2P) system that outputs fully-specified IPA transcriptions. Our approach adapts an existing diacritization model with lightweight adaptors, incurring negligible additional latency. We also contribute the ILSpeech dataset of transcribed Hebrew speech with IPA annotations, serving as a benchmark for Hebrew G2P, as training data for TTS systems, and enabling audio-to-IPA for evaluating TTS performance while capturing important phonetic details. Our results demonstrate that Phonikud G2P conversion more accurately predicts phonemes from Hebrew text compared to prior methods, and that this enables training of effective real-time Hebrew TTS models with superior speed-accuracy trade-offs. We release our code, data, and models at https: //phonikud.github.io.

comment: Project page: https://phonikud.github.io

☆ Accent-Invariant Automatic Speech Recognition via Saliency-Driven Spectrogram Masking ICASSP 2026

Mohammad Hossein Sameti, Sepehr Harfi Moridani, Ali Zarean, Hossein Sameti

Pre-trained transformer-based models have significantly advanced automatic speech recognition (ASR), yet they remain sensitive to accent and dialectal variations, resulting in elevated word error rates (WER) in linguistically diverse languages such as English and Persian. To address this challenge, we propose an accent-invariant ASR framework that integrates accent and dialect classification into the recognition pipeline. Our approach involves training a spectrogram-based classifier to capture accent-specific cues, masking the regions most influential to its predictions, and using the masked spectrograms for data augmentation. This enhances the robustness of ASR models against accent variability. We evaluate the method using both English and Persian speech. For Persian, we introduce a newly collected dataset spanning multiple regional accents, establishing the first systematic benchmark for accent variation in Persian ASR that fills a critical gap in multilingual speech research and provides a foundation for future studies on low-resource, linguistically diverse languages. Experimental results with the Whisper model demonstrate that our masking and augmentation strategy yields substantial WER reductions in both English and Persian settings, confirming the effectiveness of the approach. This research advances the development of multilingual ASR systems that are resilient to accent and dialect diversity. Code and dataset are publicly available at: https://github.com/MH-Sameti/Accent_invariant_ASR

comment: Submitted to ICASSP 2026

☆ Spatially-Augmented Sequence-to-Sequence Neural Diarization for Meetings ICASSP 2026

Li Li, Ming Cheng, Hongyu Zhang, Juan Liu, Ming Li

This paper proposes a Spatially-Augmented Sequence-to-Sequence Neural Diarization (SA-S2SND) framework, which integrates direction-of-arrival (DOA) cues estimated by SRP-DNN into the S2SND backbone. A two-stage training strategy is adopted: the model is first trained with single-channel audio and DOA features, and then further optimized with multi-channel inputs under DOA guidance. In addition, a simulated DOA generation scheme is introduced to alleviate dependence on matched multi-channel corpora. On the AliMeeting dataset, SA-S2SND consistently outperform the S2SND baseline, achieving a 7.4% relative DER reduction in the offline mode and over 19% improvement when combined with channel attention. These results demonstrate that spatial cues are highly complementary to cross-channel modeling, yielding good performance in both online and offline settings.

comment: This paper has submitted to ICASSP 2026

☆ A Study of the Removability of Speaker-Adversarial Perturbations

Liping Chen, Chenyang Guo, Kong Aik Lee, Zhen-Hua Ling, Wu Guo

Recent advancements in adversarial attacks have demonstrated their effectiveness in misleading speaker recognition models, making wrong predictions about speaker identities. On the other hand, defense techniques against speaker-adversarial attacks focus on reducing the effects of speaker-adversarial perturbations on speaker attribute extraction. These techniques do not seek to fully remove the perturbations and restore the original speech. To this end, this paper studies the removability of speaker-adversarial perturbations. Specifically, the investigation is conducted assuming various degrees of awareness of the perturbation generator across three scenarios: ignorant, semi-informed, and well-informed. Besides, we consider both the optimization-based and feedforward perturbation generation methods. Experiments conducted on the LibriSpeech dataset demonstrated that: 1) in the ignorant scenario, speaker-adversarial perturbations cannot be eliminated, although their impact on speaker attribute extraction is reduced, 2) in the semi-informed scenario, the speaker-adversarial perturbations cannot be fully removed, while those generated by the feedforward model can be considerably reduced, and 3) in the well-informed scenario, speaker-adversarial perturbations are nearly eliminated, allowing for the restoration of the original speech. Audio samples can be found in https://voiceprivacy.github.io/Perturbation-Generation-Removal/.

☆ The Speech-LLM Takes It All: A Truly Fully End-to-End Spoken Dialogue State Tracking Approach

Nizar El Ghazal, Antoine Caubrière, Valentin Vielzeuf

This paper presents a comparative study of context management strategies for end-to-end Spoken Dialog State Tracking using Speech-LLMs. We systematically evaluate traditional multimodal context (combining text history and spoken current turn), full spoken history, and compressed spoken history approaches. Our experiments on the SpokenWOZ corpus demonstrate that providing the full spoken conversation as input yields the highest performance among models of similar size, significantly surpassing prior methods. Furthermore, we show that attention-pooling-based compression of the spoken history offers a strong trade-off, maintaining competitive accuracy with reduced context size. Detailed analysis confirms that improvements stem from more effective context utilization.

☆ WildElder: A Chinese Elderly Speech Dataset from the Wild with Fine-Grained Manual Annotations

Hui Wang, Jiaming Zhou, Jiabei He, Haoqin Sun, Yong Qin

Elderly speech poses unique challenges for automatic processing due to age-related changes such as slower articulation and vocal tremors. Existing Chinese datasets are mostly recorded in controlled environments, limiting their diversity and real-world applicability. To address this gap, we present WildElder, a Mandarin elderly speech corpus collected from online videos and enriched with fine-grained manual annotations, including transcription, speaker age, gender, and accent strength. Combining the realism of in-the-wild data with expert curation, WildElder enables robust research on automatic speech recognition and speaker profiling. Experimental results reveal both the difficulties of elderly speech recognition and the potential of WildElder as a challenging new benchmark. The dataset and code are available at https://github.com/NKU-HLT/WildElder.

☆ Target speaker anonymization in multi-speaker recordings ICASSP 2026

Natalia Tomashenko, Junichi Yamagishi, Xin Wang, Yun Liu, Emmanuel Vincent

Most of the existing speaker anonymization research has focused on single-speaker audio, leading to the development of techniques and evaluation metrics optimized for such condition. This study addresses the significant challenge of speaker anonymization within multi-speaker conversational audio, specifically when only a single target speaker needs to be anonymized. This scenario is highly relevant in contexts like call centers, where customer privacy necessitates anonymizing only the customer's voice in interactions with operators. Conventional anonymization methods are often not suitable for this task. Moreover, current evaluation methodology does not allow us to accurately assess privacy protection and utility in this complex multi-speaker scenario. This work aims to bridge these gaps by exploring effective strategies for targeted speaker anonymization in conversational audio, highlighting potential problems in their development and proposing corresponding improved evaluation methodologies.

comment: Submitted to ICASSP 2026

☆ SynthVC: Leveraging Synthetic Data for End-to-End Low Latency Streaming Voice Conversion SC2025

Zhao Guo, Ziqian Ning, Guobin Ma, Lei Xie

Voice Conversion (VC) aims to modify a speaker's timbre while preserving linguistic content. While recent VC models achieve strong performance, most struggle in real-time streaming scenarios due to high latency, dependence on ASR modules, or complex speaker disentanglement, which often results in timbre leakage or degraded naturalness. We present SynthVC, a streaming end-to-end VC framework that directly learns speaker timbre transformation from synthetic parallel data generated by a pre-trained zero-shot VC model. This design eliminates the need for explicit content-speaker separation or recognition modules. Built upon a neural audio codec architecture, SynthVC supports low-latency streaming inference with high output fidelity. Experimental results show that SynthVC outperforms baseline streaming VC systems in both naturalness and speaker similarity, achieving an end-to-end latency of just 77.1 ms.

comment: Accepted by NCMMSC2025

☆ Effects of automotive microphone frequency response characteristics and noise conditions on speech and ASR quality -- an experimental evaluation

Michele Buccoli, Yu Du, Jacob Soendergaard, Simone Shawn Cazzaniga

Upon choosing microphones for automotive hands-free communication or Automatic Speech Recognition (ASR) applications, OEMs typically specify wideband, super wideband or even fullband requirements following established standard recommendations (e.g., ITU-P.1110, ITU-P.1120). In practice, it is often challenging to achieve the preferred bandwidth for an automotive microphone when considering limitations and constraints on microphone placement inside the cabin, and the automotive grade environmental robustness requirements. On the other hand, there seems to be no consensus or sufficient data on the effect of each microphone characteristic on the actual performance. As an attempt to answer this question, we used noise signals recorded in real vehicles and under various driving conditions to experimentally study the relationship between the microphones' characteristics and the final audio quality of speech communication and performance of ASR engines. We focus on how variations in microphone bandwidth and amplitude frequency response shapes affect the perceptual speech quality. The speech quality results are compared by using ETSI TS 103 281 metrics (S-MOS, N-MOS, G-MOS) and ancillary metrics such as SNR. The ASR results are evaluated with standard metrics such as Word Error Rate (WER). Findings from this study provide knowledge in the understanding of what microphone frequency response characteristics are more relevant for audio quality and choice of proper microphone specifications, particularly for automotive applications.

☆ Unsupervised lexicon learning from speech is limited by representations rather than clustering ICASSP 2026

Danel Adendorff, Simon Malan, Herman Kamper

Zero-resource word segmentation and clustering systems aim to tokenise speech into word-like units without access to text labels. Despite progress, the induced lexicons are still far from perfect. In an idealised setting with gold word boundaries, we ask whether performance is limited by the representation of word segments, or by the clustering methods that group them into word-like types. We combine a range of self-supervised speech features (continuous/discrete, frame/word-level) with different clustering methods (K-means, hierarchical, graph-based) on English and Mandarin data. The best system uses graph clustering with dynamic time warping on continuous features. Faster alternatives use graph clustering with cosine distance on averaged continuous features or edit distance on discrete unit sequences. Through controlled experiments that isolate either the representations or the clustering method, we demonstrate that representation variability across segments of the same word type -- rather than clustering -- is the primary factor limiting performance.

comment: Submitted to ICASSP 2026

☆ Impact of HRTF individualisation and head movements in a real/virtual localisation task

Vincent Martin, Lorenzo Picinali

The objective of Audio Augmented Reality (AAR) applications are to seamlessly integrate virtual sound sources within a real environment. It is critical for these applications that virtual sources are localised precisely at the intended position, and that the acoustic environments are accurately matched. One effective method for spatialising sound on headphones is through Head-Related Transfer Functions (HRTFs). These characterise how the physical features of a listener modify sound waves before they reach the eardrum. This study examines the influence of using individualised HRTFs on the localisation and the perceived realism of virtual sound sources associated with a real visual object. Participants were tasked with localising virtual and real speech sources presented via headphones and through a spherical loudspeaker array, respectively. The assessment focussed on perceived realism and sources location. All sources were associated with one of thirty real visual sources (loudspeakers) arranged in a semi-anechoic room. Various sound source renderings were compared, including single loudspeaker rendering and binaural rendering with individualised or non-individualised HRTFs. Additionally, the impact of head movements was explored: ten participants completed the same task with and without the possibility to move their head. The results showed that using individual HRTFs improved perceived realism but not localisation performance in the static scenario. Surprisingly, the opposite was observed when head movements were possible and encouraged.

☆ FLToP CTC: Frame-Level Token Pruning via Relative Threshold for Efficient and Memory-Saving Decoding on Diverse Platforms

Atul Shree, Harshith Jupuru

CTC-based ASR systems face computational and memory bottlenecks in resource-limited environments. Traditional CTC decoders, requiring up to 90% of processing time in systems (e.g., wav2vec2-large on L4 GPUs), face inefficiencies due to exhaustive token-level operations. This paper introduces Frame Level Token Pruning for Connectionist Temporal Classification (FLToP CTC), a novel decoding algorithm that employs frame-level token pruning guided by a relative threshold probability. By dynamically eliminating low-probability tokens per frame, FLToP CTC reduces compute and memory demands while maintaining negligible WER degradation. On LibriSpeech, FLToP CTC achieves a 10.5x runtime speedup and 2.78x memory reduction versus standard CTC decoders. Its simplicity enables seamless integration into CTC decoders across platforms (CPUs, GPUs, etc.). FLToP CTC addresses CTC bottlenecks, offering scalability for resource-limited environments and realtime applications, enhancing speech recognition accessibility and efficiency.

comment: 5 pages, 5 figures

☆ Emotion-Disentangled Embedding Alignment for Noise-Robust and Cross-Corpus Speech Emotion Recognition

Upasana Tiwari, Rupayan Chakraborty, Sunil Kumar Kopparapu

Effectiveness of speech emotion recognition in real-world scenarios is often hindered by noisy environments and variability across datasets. This paper introduces a two-step approach to enhance the robustness and generalization of speech emotion recognition models through improved representation learning. First, our model employs EDRL (Emotion-Disentangled Representation Learning) to extract class-specific discriminative features while preserving shared similarities across emotion categories. Next, MEA (Multiblock Embedding Alignment) refines these representations by projecting them into a joint discriminative latent subspace that maximizes covariance with the original speech input. The learned EDRL-MEA embeddings are subsequently used to train an emotion classifier using clean samples from publicly available datasets, and are evaluated on unseen noisy and cross-corpus speech samples. Improved performance under these challenging conditions demonstrates the effectiveness of the proposed method.

comment: 13 pages, 1 figure

☆ MMAudioSep: Taming Video-to-Audio Generative Model Towards Video/Text-Queried Sound Separation

Akira Takahashi, Shusuke Takahashi, Yuki Mitsufuji

We introduce MMAudioSep, a generative model for video/text-queried sound separation that is founded on a pretrained video-to-audio model. By leveraging knowledge about the relationship between video/text and audio learned through a pretrained audio generative model, we can train the model more efficiently, i.e., the model does not need to be trained from scratch. We evaluate the performance of MMAudioSep by comparing it to existing separation models, including models based on both deterministic and generative approaches, and find it is superior to the baseline models. Furthermore, we demonstrate that even after acquiring functionality for sound separation via fine-tuning, the model retains the ability for original video-to-audio generation. This highlights the potential of foundational sound generation models to be adopted for sound-related downstream tasks. Our code is available at https://github.com/sony/mmaudiosep.

comment: 4 pages, 4 figures, 2 tables

☆ O_O-VC: Synthetic Data-Driven One-to-One Alignment for Any-to-Any Voice Conversion EMNLP 2025

Huu Tuong Tu, Huan Vu, cuong tien nguyen, Dien Hy Ngo, Nguyen Thi Thu Trang

Traditional voice conversion (VC) methods typically attempt to separate speaker identity and linguistic information into distinct representations, which are then combined to reconstruct the audio. However, effectively disentangling these factors remains challenging, often leading to information loss during training. In this paper, we propose a new approach that leverages synthetic speech data generated by a high-quality, pretrained multispeaker text-to-speech (TTS) model. Specifically, synthetic data pairs that share the same linguistic content but differ in speaker identity are used as input-output pairs to train the voice conversion model. This enables the model to learn a direct mapping between source and target voices, effectively capturing speaker-specific characteristics while preserving linguistic content. Additionally, we introduce a flexible training strategy for any-to-any voice conversion that generalizes well to unseen speakers and new languages, enhancing adaptability and performance in zero-shot scenarios. Our experiments show that our proposed method achieves a 16.35% relative reduction in word error rate and a 5.91% improvement in speaker cosine similarity, outperforming several state-of-the-art methods. Voice conversion samples can be accessed at: https://oovc-emnlp-2025.github.io/

comment: EMNLP 2025

☆ Déréverbération non-supervisée de la parole par modèle hybride

Louis Bahrman, Mathieu Fontaine, Gaël Richard

This paper introduces a new training strategy to improve speech dereverberation systems in an unsupervised manner using only reverberant speech. Most existing algorithms rely on paired dry/reverberant data, which is difficult to obtain. Our approach uses limited acoustic information, like the reverberation time (RT60), to train a dereverberation system. Experimental results demonstrate that our method achieves more consistent performance across various objective metrics than the state-of-the-art.

comment: in French language

☆ DiTSinger: Scaling Singing Voice Synthesis with Diffusion Transformer and Implicit Alignment

Zongcai Du, Guilin Deng, Xiaofeng Guo, Xin Gao, Linke Li, Kaichang Cheng, Fubo Han, Siyu Yang, Peng Liu, Pan Zhong, Qiang Fu

Recent progress in diffusion-based Singing Voice Synthesis (SVS) demonstrates strong expressiveness but remains limited by data scarcity and model scalability. We introduce a two-stage pipeline: a compact seed set of human-sung recordings is constructed by pairing fixed melodies with diverse LLM-generated lyrics, and melody-specific models are trained to synthesize over 500 hours of high-quality Chinese singing data. Building on this corpus, we propose DiTSinger, a Diffusion Transformer with RoPE and qk-norm, systematically scaled in depth, width, and resolution for enhanced fidelity. Furthermore, we design an implicit alignment mechanism that obviates phoneme-level duration labels by constraining phoneme-to-acoustic attention within character-level spans, thereby improving robustness under noisy or uncertain alignments. Extensive experiments validate that our approach enables scalable, alignment-free, and high-fidelity SVS.

comment: under review

☆ VM-UNSSOR: Unsupervised Neural Speech Separation Enhanced by Higher-SNR Virtual Microphone Arrays

Shulin He, Zhong-Qiu Wang

Blind speech separation (BSS) aims to recover multiple speech sources from multi-channel, multi-speaker mixtures under unknown array geometry and room impulse responses. In unsupervised setup where clean target speech is not available for model training, UNSSOR proposes a mixture consistency (MC) loss for training deep neural networks (DNN) on over-determined training mixtures to realize unsupervised speech separation. However, when the number of microphones of the training mixtures decreases, the MC constraint weakens and the separation performance falls dramatically. To address this, we propose VM-UNSSOR, augmenting the observed training mixture signals recorded by a limited number of microphones with several higher-SNR virtual-microphone (VM) signals, which are obtained by applying linear spatial demixers (such as IVA and spatial clustering) to the observed training mixtures. As linear projections of the observed mixtures, the virtual-microphone signals can typically increase the SNR of each source and can be leveraged to compute extra MC losses to improve UNSSOR and address the frequency permutation problem in UNSSOR. On the SMS-WSJ dataset, in the over-determined six-microphone, two-speaker separation setup, VM-UNSSOR reaches 17.1 dB SI-SDR, while UNSSOR only obtains 14.7 dB; and in the determined two-microphone, two-speaker case, UNSSOR collapses to -2.7 dB SI-SDR, while VM-UNSSOR achieves 10.7 dB.

☆ ControlAudio: Tackling Text-Guided, Timing-Indicated and Intelligible Audio Generation via Progressive Diffusion Modeling

Yuxuan Jiang, Zehua Chen, Zeqian Ju, Yusheng Dai, Weibei Dou, Jun Zhu

Text-to-audio (TTA) generation with fine-grained control signals, e.g., precise timing control or intelligible speech content, has been explored in recent works. However, constrained by data scarcity, their generation performance at scale is still compromised. In this study, we recast controllable TTA generation as a multi-task learning problem and introduce a progressive diffusion modeling approach, ControlAudio. Our method adeptly fits distributions conditioned on more fine-grained information, including text, timing, and phoneme features, through a step-by-step strategy. First, we propose a data construction method spanning both annotation and simulation, augmenting condition information in the sequence of text, timing, and phoneme. Second, at the model training stage, we pretrain a diffusion transformer (DiT) on large-scale text-audio pairs, achieving scalable TTA generation, and then incrementally integrate the timing and phoneme features with unified semantic representations, expanding controllability. Finally, at the inference stage, we propose progressively guided generation, which sequentially emphasizes more fine-grained information, aligning inherently with the coarse-to-fine sampling nature of DiT. Extensive experiments show that ControlAudio achieves state-of-the-art performance in terms of temporal accuracy and speech clarity, significantly outperforming existing methods on both objective and subjective evaluations. Demo samples are available at: https://control-audio.github.io/Control-Audio.

comment: 18 pages, 8 tables, 5 figures

☆ Serial-Parallel Dual-Path Architecture for Speaking Style Recognition SC2025

Guojian Li, Qijie Shao, Zhixian Zhao, Shuiyuan Wang, Zhonghua Fu, Lei Xie

Speaking Style Recognition (SSR) identifies a speaker's speaking style characteristics from speech. Existing style recognition approaches primarily rely on linguistic information, with limited integration of acoustic information, which restricts recognition accuracy improvements. The fusion of acoustic and linguistic modalities offers significant potential to enhance recognition performance. In this paper, we propose a novel serial-parallel dual-path architecture for SSR that leverages acoustic-linguistic bimodal information. The serial path follows the ASR+STYLE serial paradigm, reflecting a sequential temporal dependency, while the parallel path integrates our designed Acoustic-Linguistic Similarity Module (ALSM) to facilitate cross-modal interaction with temporal simultaneity. Compared to the existing SSR baseline -- the OSUM model, our approach reduces parameter size by 88.4% and achieves a 30.3% improvement in SSR accuracy for eight styles on the test set.

comment: Accepted by NCMMSC2025

☆ Production and Manufacturing of 3D Printed Acoustic Guitars

Timothy Tran, William Schiesser

This research investigates the feasibility of producing affordable, functional acoustic guitars using 3D printing, with a focus on producing structural designs with proper tonal performance. Conducted in collaboration with William Schiesser, the study uses a classical guitar model, chosen for its lower string tension, to evaluate the tonal characteristics of a 3D-printed prototype made from polylactic acid (PLA). Due to the build plate size constraints of the Prusa Mark 4 printer, the guitar body was divided into multiple sections joined with press-fit tolerances and minimal cyanoacrylate adhesive. CAD modeling in Fusion 360 ensured dimensional accuracy in press-fit connections and the overall assembly. Following assembly, the guitar was strung with nylon strings and tested using Audacity software to compare recorded frequencies and notes with standard reference values. Results showed large deviations in lower string frequencies, likely caused by the material choice utilized in printing. Accurate pitches were reached with all strings despite frequency differences through tuning, demonstrating that PLA and modern manufacturing methods can produce affordable, playable acoustic guitars despite inevitable challenges. Further research may investigate alternative plastics for superior frequency matching. This approach holds significant potential for expanding access to quality instruments while reducing reliance on endangered tonewoods, thereby encouraging both sustainable instrument production and increased musical participation. This also creates opportunities for disadvantaged communities where access to musical instruments remains a challenge. Keywords: Luthiery, Stereolithography, 3D-Print, Guitar Making

♻ ☆ Iola Walker: A Mobile Footfall Detection System for Music Composition

William B. James

This outing is part of a larger music technology research project. The objective is to find a method for materially enhancing music using hardware and software. There is a strong likelihood that there exists a new medium for experiencing music via a wearable device that ordinary listeners prefer over the current state of the art. If such a medium is discovered, it is a step towards altruistic, prosocial reform in the music industry. A new playback system infrastructure has a chance to soothe some of the societal problems tied to the larger entertainment industry ecosystem. Iola walker is a music playback system that allows musicians to compose music that changes in accordance with the listener's gait. Artifacts are available here: https://github.com/willbjames/iolawalker

♻ ☆ Direction Estimation of Sound Sources Using Microphone Arrays and Signal Strength ICSE

Mahdi Ali Pour, Zahra Habibzadeh

Sound-tracking refers to the process of determining the direction from which a sound originates, making it a fundamental component of sound source localization. This capability is essential in a variety of applications, including security systems, acoustic monitoring, and speaker tracking, where accurately identifying the direction of a sound source enables real-time responses, efficient resource allocation, and improved situational awareness. While sound-tracking is closely related to localization, it specifically focuses on identifying the direction of the sound source rather than estimating its exact position in space. Despite its utility, sound-tracking systems face several challenges, such as maintaining directional accuracy and precision, along with the need for sophisticated hardware configurations and complex signal processing algorithms. This paper presents a sound-tracking method using three electret microphones. We estimate the direction of a sound source using a lightweight method that analyzes signals from three strategically placed microphones. By comparing the average power of the received signals, the system infers the most probable direction of the sound. The results indicate that the power level from each microphone effectively determines the sound source direction. Our system employs a straightforward and cost-effective hardware design, ensuring simplicity and affordability in implementation. It achieves a localization error of less than 6 degrees and a precision of 98%. Additionally, its effortless integration with various systems makes it versatile and adaptable. Consequently, this technique presents a robust and reliable solution for sound-tracking and localization, with potential applications spanning diverse domains such as security systems, smart homes, and acoustic monitoring.

comment: Accepted to the 32nd International Conference on Systems Engineering (ICSEng'2025)

♻ ☆ Phonikud: Hebrew Grapheme-to-Phoneme Conversion for Real-Time Text-to-Speech

Yakov Kolani, Maxim Melichov, Cobi Calev, Morris Alper

Real-time text-to-speech (TTS) for Modern Hebrew is challenging due to the language's orthographic complexity. Existing solutions ignore crucial phonetic features such as stress that remain underspecified even when vowel marks are added. To address these limitations, we introduce Phonikud, a lightweight, open-source Hebrew grapheme-to-phoneme (G2P) system that outputs fully-specified IPA transcriptions. Our approach adapts an existing diacritization model with lightweight adaptors, incurring negligible additional latency. We also contribute the ILSpeech dataset of transcribed Hebrew speech with IPA annotations, serving as a benchmark for Hebrew G2P, as training data for TTS systems, and enabling audio-to-IPA for evaluating TTS performance while capturing important phonetic details. Our results demonstrate that Phonikud G2P conversion more accurately predicts phonemes from Hebrew text compared to prior methods, and that this enables training of effective real-time Hebrew TTS models with superior speed-accuracy trade-offs. We release our code, data, and models at https: //phonikud.github.io.

comment: Project page: https://phonikud.github.io

☆ MeanVC: Lightweight and Streaming Zero-Shot Voice Conversion via Mean Flows

Guobin Ma, Jixun Yao, Ziqian Ning, Yuepeng Jiang, Lingxin Xiong, Lei Xie, Pengcheng Zhu

Zero-shot voice conversion (VC) aims to transfer timbre from a source speaker to any unseen target speaker while preserving linguistic content. Growing application scenarios demand models with streaming inference capabilities. This has created a pressing need for models that are simultaneously fast, lightweight, and high-fidelity. However, existing streaming methods typically rely on either autoregressive (AR) or non-autoregressive (NAR) frameworks, which either require large parameter sizes to achieve strong performance or struggle to generalize to unseen speakers. In this study, we propose MeanVC, a lightweight and streaming zero-shot VC approach. MeanVC introduces a diffusion transformer with a chunk-wise autoregressive denoising strategy, combining the strengths of both AR and NAR paradigms for efficient streaming processing. By introducing mean flows, MeanVC regresses the average velocity field during training, enabling zero-shot VC with superior speech quality and speaker similarity in a single sampling step by directly mapping from the start to the endpoint of the flow trajectory. Additionally, we incorporate diffusion adversarial post-training to mitigate over-smoothing and further enhance speech quality. Experimental results demonstrate that MeanVC significantly outperforms existing zero-shot streaming VC systems, achieving superior conversion quality with higher efficiency and significantly fewer parameters. Audio demos and code are publicly available at https://aslp-lab.github.io/MeanVC.

☆ Leveraging Whisper Embeddings for Audio-based Lyrics Matching

Eleonora Mancini, Joan Serrà, Paolo Torroni, Yuki Mitsufuji

Audio-based lyrics matching can be an appealing alternative to other content-based retrieval approaches, but existing methods often suffer from limited reproducibility and inconsistent baselines. In this work, we introduce WEALY, a fully reproducible pipeline that leverages Whisper decoder embeddings for lyrics matching tasks. WEALY establishes robust and transparent baselines, while also exploring multimodal extensions that integrate textual and acoustic features. Through extensive experiments on standard datasets, we demonstrate that WEALY achieves a performance comparable to state-of-the-art methods that lack reproducibility. In addition, we provide ablation studies and analyses on language robustness, loss functions, and embedding strategies. This work contributes a reliable benchmark for future research, and underscores the potential of speech technologies for music information retrieval tasks.

☆ Detecting and Mitigating Insertion Hallucination in Video-to-Audio Generation

Liyang Chen, Hongkai Chen, Yujun Cai, Sifan Li, Qingwen Ye, Yiwei Wang

Video-to-Audio generation has made remarkable strides in automatically synthesizing sound for video. However, existing evaluation metrics, which focus on semantic and temporal alignment, overlook a critical failure mode: models often generate acoustic events, particularly speech and music, that have no corresponding visual source. We term this phenomenon Insertion Hallucination and identify it as a systemic risk driven by dataset biases, such as the prevalence of off-screen sounds, that remains completely undetected by current metrics. To address this challenge, we first develop a systematic evaluation framework that employs a majority-voting ensemble of multiple audio event detectors. We also introduce two novel metrics to quantify the prevalence and severity of this issue: IH@vid (the fraction of videos with hallucinations) and IH@dur (the fraction of hallucinated duration). Building on this, we propose Posterior Feature Correction, a novel training-free inference-time method that mitigates IH. PFC operates in a two-pass process: it first generates an initial audio output to detect hallucinated segments, and then regenerates the audio after masking the corresponding video features at those timestamps. Experiments on several mainstream V2A benchmarks first reveal that state-of-the-art models suffer from severe IH. In contrast, our PFC method reduces both the prevalence and duration of hallucinations by over 50\% on average, without degrading, and in some cases even improving, conventional metrics for audio quality and temporal synchronization. Our work is the first to formally define, systematically measure, and effectively mitigate Insertion Hallucination, paving the way for more reliable and faithful V2A models.

☆ Attribution-by-design: Ensuring Inference-Time Provenance in Generative Music Systems

Fabio Morreale, Wiebke Hutiri, Joan Serrà, Alice Xiang, Yuki Mitsufuji

The rise of AI-generated music is diluting royalty pools and revealing structural flaws in existing remuneration frameworks, challenging the well-established artist compensation systems in the music industry. Existing compensation solutions, such as piecemeal licensing agreements, lack scalability and technical rigour, while current data attribution mechanisms provide only uncertain estimates and are rarely implemented in practice. This paper introduces a framework for a generative music infrastructure centred on direct attribution, transparent royalty distribution, and granular control for artists and rights' holders. We distinguish ontologically between the training set and the inference set, which allows us to propose two complementary forms of attribution: training-time attribution and inference-time attribution. We here favour inference-time attribution, as it enables direct, verifiable compensation whenever an artist's catalogue is used to condition a generated output. Besides, users benefit from the ability to condition generations on specific songs and receive transparent information about attribution and permitted usage. Our approach offers an ethical and practical solution to the pressing need for robust compensation mechanisms in the era of AI-generated music, ensuring that provenance and fairness are embedded at the core of generative systems.

☆ Personality-Enhanced Multimodal Depression Detection in the Elderly

Honghong Wang, Jing Deng, Rong Zheng

This paper presents our solution to the Multimodal Personality-aware Depression Detection (MPDD) challenge at ACM MM 2025. We propose a multimodal depression detection model in the Elderly that incorporates personality characteristics. We introduce a multi-feature fusion approach based on a co-attention mechanism to effectively integrate LLDs, MFCCs, and Wav2Vec features in the audio modality. For the video modality, we combine representations extracted from OpenFace, ResNet, and DenseNet to construct a comprehensive visual feature set. Recognizing the critical role of personality in depression detection, we design an interaction module that captures the relationships between personality traits and multimodal features. Experimental results from the MPDD Elderly Depression Detection track demonstrate that our method significantly enhances performance, providing valuable insights for future research in multimodal depression detection among elderly populations.

comment: 6 pages,2 figures,accepted by ACM Multimedia Asia 2025

☆ IntMeanFlow: Few-step Speech Generation with Integral Velocity Distillation

Wei Wang, Rong Cao, Yi Guo, Zhengyang Chen, Kuan Chen, Yuanyuan Huo

Flow-based generative models have greatly improved text-to-speech (TTS) synthesis quality, but inference speed remains limited by the iterative sampling process and multiple function evaluations (NFE). The recent MeanFlow model accelerates generation by modeling average velocity instead of instantaneous velocity. However, its direct application to TTS encounters challenges, including GPU memory overhead from Jacobian-vector products (JVP) and training instability due to self-bootstrap processes. To address these issues, we introduce IntMeanFlow, a framework for few-step speech generation with integral velocity distillation. By approximating average velocity with the teacher's instantaneous velocity over a temporal interval, IntMeanFlow eliminates the need for JVPs and self-bootstrap, improving stability and reducing GPU memory usage. We also propose the Optimal Step Sampling Search (O3S) algorithm, which identifies the model-specific optimal sampling steps, improving speech synthesis without additional inference overhead. Experiments show that IntMeanFlow achieves 1-NFE inference for token-to-spectrogram and 3-NFE for text-to-spectrogram tasks while maintaining high-quality synthesis. Demo samples are available at https://vvwangvv.github.io/intmeanflow.

☆ ACMID: Automatic Curation of Musical Instrument Dataset for 7-Stem Music Source Separation

Ji Yu, Yang shuo, Xu Yuetonghui, Liu Mengmei, Ji Qiang, Han Zerui

Most current music source separation (MSS) methods rely on supervised learning, limited by training data quan- tity and quality. Though web-crawling can bring abundant data, platform-level track labeling often causes metadata mismatches, impeding accurate "audio-label" pair acquisi- tion. To address this, we present ACMID: a dataset for MSS generated through web crawling of extensive raw data, fol- lowed by automatic cleaning via an instrument classifier built on a pre-trained audio encoder that filters and aggregates clean segments of target instruments from the crawled tracks, resulting in the refined ACMID-Cleaned dataset. Leverag- ing abundant data, we expand the conventional classifica- tion from 4-stem (Vocal/Bass/Drums/Others) to 7-stem (Pi- ano/Drums/Bass/Acoustic Guitar/Electric Guitar/Strings/Wind- Brass), enabling high granularity MSS systems. Experiments on SOTA MSS model demonstrates two key results: (i) MSS model trained with ACMID-Cleaned achieved a 2.39dB improvement in SDR performance compared to that with ACMID-Uncleaned, demostrating the effectiveness of our data cleaning procedure; (ii) incorporating ACMID-Cleaned to training enhances MSS model's average performance by 1.16dB, confirming the value of our dataset. Our data crawl- ing code, cleaning model code and weights are available at: https://github.com/scottishfold0621/ACMID.

☆ IsoSignVid2Aud: Sign Language Video to Audio Conversion without Text Intermediaries

Harsh Kavediya, Vighnesh Nayak, Bheeshm Sharma, Balamurugan Palaniappan

Sign language to spoken language audio translation is important to connect the hearing- and speech-challenged humans with others. We consider sign language videos with isolated sign sequences rather than continuous grammatical signing. Such videos are useful in educational applications and sign prompt interfaces. Towards this, we propose IsoSignVid2Aud, a novel end-to-end framework that translates sign language videos with a sequence of possibly non-grammatic continuous signs to speech without requiring intermediate text representation, providing immediate communication benefits while avoiding the latency and cascading errors inherent in multi-stage translation systems. Our approach combines an I3D-based feature extraction module with a specialized feature transformation network and an audio generation pipeline, utilizing a novel Non-Maximal Suppression (NMS) algorithm for the temporal detection of signs in non-grammatic continuous sequences. Experimental results demonstrate competitive performance on ASL-Citizen-1500 and WLASL-100 datasets with Top-1 accuracies of 72.01\% and 78.67\%, respectively, and audio quality metrics (PESQ: 2.67, STOI: 0.73) indicating intelligible speech output. Code is available at: https://github.com/BheeshmSharma/IsoSignVid2Aud_AIMLsystems-2025.

comment: Accepted in AIML-Systems-2025

☆ ACMID: Automatic Curation of Musical Instrument Dataset for 7-Stem Music Source Separation

Ji Yu, Yang shuo, Xu Yuetonghui, Liu Mengmei, Ji Qiang, Han Zerui

Most current music source separation (MSS) methods rely on supervised learning, limited by training data quantity and quality. Though web-crawling can bring abundant data, platform-level track labeling often causes metadata mismatches, impeding accurate "audio-label" pair acquisition. To address this, we present ACMID: a dataset for MSS generated through web crawling of extensive raw data, followed by automatic cleaning via an instrument classifier built on a pre-trained audio encoder that filters and aggregates clean segments of target instruments from the crawled tracks, resulting in the refined ACMID-Cleaned dataset. Leveraging abundant data, we expand the conventional classification from 4-stem (Vocal/Bass/Drums/Others) to 7-stem (Piano/Drums/Bass/Acoustic Guitar/Electric Guitar/Strings/Wind-Brass), enabling high granularity MSS systems. Experiments on SOTA MSS model demonstrates two key results: (i) MSS model trained with ACMID-Cleaned achieved a 2.39dB improvement in SDR performance compared to that with ACMID-Uncleaned, demostrating the effectiveness of our data cleaning procedure; (ii) incorporating ACMID-Cleaned to training enhances MSS model's average performance by 1.16dB, confirming the value of our dataset. Our data crawling code, cleaning model code and weights are available at: https://github.com/scottishfold0621/ACMID.

☆ Audible Networks: Deconstructing and Manipulating Sounds with Deep Non-Negative Autoencoders

Juan José Burred, Carmine-Emanuele Cella

We propose the use of Non-Negative Autoencoders (NAEs) for sound deconstruction and user-guided manipulation of sounds for creative purposes. NAEs offer a versatile and scalable extension of traditional Non-Negative Matrix Factorization (NMF)-based approaches for interpretable audio decomposition. By enforcing non-negativity constraints through projected gradient descent, we obtain decompositions where internal weights and activations can be directly interpreted as spectral shapes and temporal envelopes, and where components can themselves be listened to as individual sound events. In particular, multi-layer Deep NAE architectures enable hierarchical representations with an adjustable level of granularity, allowing sounds to be deconstructed at multiple levels of abstraction: from high-level note envelopes down to fine-grained spectral details. This framework enables a wide new range of expressive, controllable, and randomized sound transformations. We introduce novel manipulation operations including cross-component and cross-layer synthesis, hierarchical deconstructions, and several randomization strategies that control timbre and event density. Through visualizations and resynthesis of practical examples, we demonstrate how NAEs can serve as flexible and interpretable tools for object-based sound editing.

☆ Beyond Discrete Categories: Multi-Task Valence-Arousal Modeling for Pet Vocalization Analysis

Junyao Huang, Rumin Situ

Traditional pet emotion recognition from vocalizations, based on discrete classification, struggles with ambiguity and capturing intensity variations. We propose a continuous Valence-Arousal (VA) model that represents emotions in a two-dimensional space. Our method uses an automatic VA label generation algorithm, enabling large-scale annotation of 42,553 pet vocalization samples. A multi-task learning framework jointly trains VA regression with auxiliary tasks (emotion, body size, gender) to enhance prediction by improving feature learning. Our Audio Transformer model achieves a validation Valence Pearson correlation of r = 0.9024 and an Arousal r = 0.7155, effectively resolving confusion between discrete categories like "territorial" and "happy." This work introduces the first continuous VA framework for pet vocalization analysis, offering a more expressive representation for human-pet interaction, veterinary diagnostics, and behavioral training. The approach shows strong potential for deployment in consumer products like AI pet emotion translators.

comment: 24 pages, 6 figures, 4 tables. First continuous VA framework for pet vocalization analysis with 42,553 samples

♻ ☆ I$^2$RF-TFCKD: Intra-Inter Representation Fusion with Time-Frequency Calibration Knowledge Distillation for Speech Enhancement

Jiaming Cheng, Ruiyu Liang, Ye Ni, Chao Xu, Jing Li, Wei Zhou, Rui Liu, Björn W. Schuller, Xiaoshuai Hao

In this paper, we propose an intra-inter representation fusion knowledge distillation (KD) framework with time-frequency calibration (I$^2$RF-TFCKD) for SE, which achieves distillation through the fusion of multi-layer teacher-student feature flows. Different from previous distillation strategies for SE, the proposed framework fully utilizes the time-frequency differential information of speech while promoting global knowledge flow. Firstly, we construct a collaborative distillation paradigm for intra-set and inter-set correlations. Within a correlated set, multi-layer teacher-student features are pairwise matched for calibrated distillation. Subsequently, we generate representative features from each correlated set through residual fusion to form the fused feature set that enables inter-set knowledge interaction. Secondly, we propose a multi-layer interactive distillation based on dual-stream time-frequency cross-calibration, which calculates the teacher-student similarity calibration weights in the time and frequency domains respectively and performs cross-weighting, thus enabling refined allocation of distillation contributions across different layers according to speech characteristics. The proposed distillation strategy is applied to the dual-path dilated convolutional recurrent network (DPDCRN) that ranked first in the SE track of the L3DAS23 challenge. To evaluate the effectiveness of I$^2$RF-TFCKD, we conduct experiments on both single-channel and multi-channel SE datasets. Objective evaluations demonstrate that the proposed KD strategy consistently and effectively improves the performance of the low-complexity student model and outperforms other distillation schemes.

comment: submitted to Information Fusion

♻ ☆ Provable Speech Attributes Conversion via Latent Independence

Jonathan Svirsky, Ofir Lindenbaum, Uri Shaham

While signal conversion and disentangled representation learning have shown promise for manipulating data attributes across domains such as audio, image, and multimodal generation, existing approaches, especially for speech style conversion, are largely empirical and lack rigorous theoretical foundations to guarantee reliable and interpretable control. In this work, we propose a general framework for speech attribute conversion, accompanied by theoretical analysis and guarantees under reasonable assumptions. Our framework builds on a non-probabilistic autoencoder architecture with an independence constraint between the predicted latent variable and the target controllable variable. This design ensures a consistent signal transformation, conditioned on an observed style variable, while preserving the original content and modifying the desired attribute. We further demonstrate the versatility of our method by evaluating it on speech styles, including speaker identity and emotion. Quantitative evaluations confirm the effectiveness and generality of the proposed approach.

♻ ☆ STOPA: A Database of Systematic VariaTion Of DeePfake Audio for Open-Set Source Tracing and Attribution

Anton Firc, Manasi Chhibber, Jagabandhu Mishra, Vishwanath Pratap Singh, Tomi Kinnunen, Kamil Malinka

A key research area in deepfake speech detection is source tracing - determining the origin of synthesised utterances. The approaches may involve identifying the acoustic model (AM), vocoder model (VM), or other generation-specific parameters. However, progress is limited by the lack of a dedicated, systematically curated dataset. To address this, we introduce STOPA, a systematically varied and metadata-rich dataset for deepfake speech source tracing, covering 8 AMs, 6 VMs, and diverse parameter settings across 700k samples from 13 distinct synthesisers. Unlike existing datasets, which often feature limited variation or sparse metadata, STOPA provides a systematically controlled framework covering a broader range of generative factors, such as the choice of the vocoder model, acoustic model, or pretrained weights, ensuring higher attribution reliability. This control improves attribution accuracy, aiding forensic analysis, deepfake detection, and generative model transparency.

comment: Published at Interspeech 2025 conference

♻ ☆ Audio-Conditioned Diffusion LLMs for ASR and Deliberation Processing

Mengqi Wang, Zhan Liu, Zengrui Jin, Guangzhi Sun, Chao Zhang, Philip C. Woodland

Diffusion-based large language models (DLLMs) have recently attracted growing interest as an alternative to autoregressive decoders. In this work, we present an empirical study on using the diffusion-based large language model LLaDA for automatic speech recognition (ASR). We first investigate its use as an external deliberation-based processing module for Whisper-LLaMA transcripts. By leveraging the bidirectional attention and denoising capabilities of LLaDA, we explore random masking, low-confidence masking, and semi-autoregressive strategies, showing that Whisper-LLaDA substantially reduces WER compared with the baseline. On LibriSpeech, the best cascade system achieves 2.25%/4.94% WER on test-clean/test-other, representing a 12.3% relative improvement over the Whisper-LLaMA baseline on the test-other split. In contrast, a plain-text LLaDA without acoustic features fails to improve accuracy, highlighting the importance of audio-conditioned embeddings. We further evaluate Whisper-LLaDA as a standalone decoder for ASR with diffusion-based and semi-autoregressive decoding. Most experimental configurations achieve faster inference than the Whisper-LLaMA baseline, although recognition accuracy is slightly lower. These findings offer an empirical view of diffusion-based LLMs for ASR and point to promising directions for improvements.

♻ ☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026

Vaibhav Srivastav, Steven Zheng, Eric Bezzam, Eustache Le Bihan, Nithin Koluguri, Piotr Żelasko, Somshubra Majumdar, Adel Moumen, Sanchit Gandhi

Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard ; Code: https://github.com/huggingface/open_asr_leaderboard

♻ ☆ SeamlessEdit: Background Noise Aware Zero-Shot Speech Editing with in-Context Enhancement

Kuan-Yu Chen, Jeng-Lin Li, Jian-Jiun Ding

With the fast development of zero-shot text-to-speech technologies, it is possible to generate high-quality speech signals that are indistinguishable from the real ones. Speech editing, including speech insertion and replacement, appeals to researchers due to its potential applications. However, existing studies only considered clean speech scenarios. In real-world applications, the existence of environmental noise could significantly degrade the quality of generation. In this study, we propose a noise-resilient speech editing framework, SeamlessEdit, for noisy speech editing. SeamlessEdit adopts a frequency-band-aware noise suppression module and an in-content refinement strategy. It can well address the scenario where the frequency bands of voice and background noise are not separated. The proposed SeamlessEdit framework outperforms state-of-the-art approaches in multiple quantitative and qualitative evaluations.

comment: 5 pages, 3 figures

☆ MeanVC: Lightweight and Streaming Zero-Shot Voice Conversion via Mean Flows

Guobin Ma, Jixun Yao, Ziqian Ning, Yuepeng Jiang, Lingxin Xiong, Lei Xie, Pengcheng Zhu

Zero-shot voice conversion (VC) aims to transfer timbre from a source speaker to any unseen target speaker while preserving linguistic content. Growing application scenarios demand models with streaming inference capabilities. This has created a pressing need for models that are simultaneously fast, lightweight, and high-fidelity. However, existing streaming methods typically rely on either autoregressive (AR) or non-autoregressive (NAR) frameworks, which either require large parameter sizes to achieve strong performance or struggle to generalize to unseen speakers. In this study, we propose MeanVC, a lightweight and streaming zero-shot VC approach. MeanVC introduces a diffusion transformer with a chunk-wise autoregressive denoising strategy, combining the strengths of both AR and NAR paradigms for efficient streaming processing. By introducing mean flows, MeanVC regresses the average velocity field during training, enabling zero-shot VC with superior speech quality and speaker similarity in a single sampling step by directly mapping from the start to the endpoint of the flow trajectory. Additionally, we incorporate diffusion adversarial post-training to mitigate over-smoothing and further enhance speech quality. Experimental results demonstrate that MeanVC significantly outperforms existing zero-shot streaming VC systems, achieving superior conversion quality with higher efficiency and significantly fewer parameters. Audio demos and code are publicly available at https://aslp-lab.github.io/MeanVC.

☆ DialoSpeech: Dual-Speaker Dialogue Generation with LLM and Flow Matching

Hanke Xie, Dake Guo, Chengyou Wang, Yue Li, Wenjie Tian, Xinfa Zhu, Xinsheng Wang, Xiulin Li, Guanqiong Miao, Bo Liu, Lei Xie

Recent advances in text-to-speech (TTS) synthesis, particularly those leveraging large language models (LLMs), have significantly improved expressiveness and naturalness. However, generating human-like, interactive dialogue speech remains challenging. Current systems face limitations due to the scarcity of dual-track data and difficulties in achieving naturalness, contextual coherence, and interactional dynamics, such as turn-taking, overlapping speech, and speaker consistency, in multi-turn conversations. To address these challenges, we propose DialoSpeech, a dual-track architecture combining a large language model with Chunked Flow Matching for expressive, human-like dialogue speech synthesis. DialoSpeech generates natural multi-turn conversations with coherent speaker turns and natural overlaps, supporting both Chinese and English and cross-lingual speech synthesis. We introduce a data processing pipeline to construct dual-track dialogue datasets, facilitating scalable training and experimental validation. Experiments show that our model outperforms baselines, offering a solution for generating human-like spoken dialogues. Audio samples are available at https://tiamojames.github.io/DialoSpeech

☆ Leveraging Whisper Embeddings for Audio-based Lyrics Matching

Eleonora Mancini, Joan Serrà, Paolo Torroni, Yuki Mitsufuji

Audio-based lyrics matching can be an appealing alternative to other content-based retrieval approaches, but existing methods often suffer from limited reproducibility and inconsistent baselines. In this work, we introduce WEALY, a fully reproducible pipeline that leverages Whisper decoder embeddings for lyrics matching tasks. WEALY establishes robust and transparent baselines, while also exploring multimodal extensions that integrate textual and acoustic features. Through extensive experiments on standard datasets, we demonstrate that WEALY achieves a performance comparable to state-of-the-art methods that lack reproducibility. In addition, we provide ablation studies and analyses on language robustness, loss functions, and embedding strategies. This work contributes a reliable benchmark for future research, and underscores the potential of speech technologies for music information retrieval tasks.

☆ Pseudo2Real: Task Arithmetic for Pseudo-Label Correction in Automatic Speech Recognition

Yi-Cheng Lin, Yu-Hsuan Li Liang, Hsuan Su, Tzu-Quan Lin, Shang-Tse Chen, Yun-Nung Chen, Hung-yi Lee

Robust ASR under domain shift is crucial because real-world systems encounter unseen accents and domains with limited labeled data. Although pseudo-labeling offers a practical workaround, it often introduces systematic, accent-specific errors that filtering fails to fix. We ask: How can we correct these recurring biases without target ground truth? We propose a simple parameter-space correction: in a source domain containing both real and pseudo-labeled data, two ASR models are fine-tuned from the same initialization, one on ground-truth labels and the other on pseudo-labels, and their weight difference forms a correction vector that captures pseudo-label biases. When applied to a pseudo-labeled target model, this vector enhances recognition, achieving up to a 35% relative Word Error Rate (WER) reduction on AfriSpeech-200 across ten African accents with the Whisper tiny model.

☆ Personality-Enhanced Multimodal Depression Detection in the Elderly

Honghong Wang, Jing Deng, Rong Zheng

This paper presents our solution to the Multimodal Personality-aware Depression Detection (MPDD) challenge at ACM MM 2025. We propose a multimodal depression detection model in the Elderly that incorporates personality characteristics. We introduce a multi-feature fusion approach based on a co-attention mechanism to effectively integrate LLDs, MFCCs, and Wav2Vec features in the audio modality. For the video modality, we combine representations extracted from OpenFace, ResNet, and DenseNet to construct a comprehensive visual feature set. Recognizing the critical role of personality in depression detection, we design an interaction module that captures the relationships between personality traits and multimodal features. Experimental results from the MPDD Elderly Depression Detection track demonstrate that our method significantly enhances performance, providing valuable insights for future research in multimodal depression detection among elderly populations.

comment: 6 pages,2 figures,accepted by ACM Multimedia Asia 2025

☆ Bloodroot: When Watermarking Turns Poisonous For Stealthy Backdoor

Kuan-Yu Chen, Yi-Cheng Lin, Jeng-Lin Li, Jian-Jiun Ding

Backdoor data poisoning is a crucial technique for ownership protection and defending against malicious attacks. Embedding hidden triggers in training data can manipulate model outputs, enabling provenance verification, and deterring unauthorized use. However, current audio backdoor methods are suboptimal, as poisoned audio often exhibits degraded perceptual quality, which is noticeable to human listeners. This work explores the intrinsic stealthiness and effectiveness of audio watermarking in achieving successful poisoning. We propose a novel Watermark-as-Trigger concept, integrated into the Bloodroot backdoor framework via adversarial LoRA fine-tuning, which enhances perceptual quality while achieving a much higher trigger success rate and clean-sample accuracy. Experiments on speech recognition (SR) and speaker identification (SID) datasets show that watermark-based poisoning remains effective under acoustic filtering and model pruning. The proposed Bloodroot backdoor framework not only secures data-to-model ownership, but also well reveals the risk of adversarial misuse.

comment: 5 pages, 3 figures

☆ Guitar Tone Morphing by Diffusion-based Model

Kuan-Yu Chen, Kuan-Lin Chen, Yu-Chieh Yu, Jian-Jiun Ding

In Music Information Retrieval (MIR), modeling and transforming the tone of musical instruments, particularly electric guitars, has gained increasing attention due to the richness of the instrument tone and the flexibility of expression. Tone morphing enables smooth transitions between different guitar sounds, giving musicians greater freedom to explore new textures and personalize their performances. This study explores learning-based approaches for guitar tone morphing, beginning with LoRA fine-tuning to improve the model performance on limited data. Moreover, we introduce a simpler method, named spherical interpolation using Music2Latent. It yields significantly better results than the more complex fine-tuning approach. Experiments show that the proposed architecture generates smoother and more natural tone transitions, making it a practical and efficient tool for music production and real-time audio effects.

comment: 5 pages

☆ ACMID: Automatic Curation of Musical Instrument Dataset for 7-Stem Music Source Separation

Ji Yu, Yang shuo, Xu Yuetonghui, Liu Mengmei, Ji Qiang, Han Zerui

Most current music source separation (MSS) methods rely on supervised learning, limited by training data quan- tity and quality. Though web-crawling can bring abundant data, platform-level track labeling often causes metadata mismatches, impeding accurate "audio-label" pair acquisi- tion. To address this, we present ACMID: a dataset for MSS generated through web crawling of extensive raw data, fol- lowed by automatic cleaning via an instrument classifier built on a pre-trained audio encoder that filters and aggregates clean segments of target instruments from the crawled tracks, resulting in the refined ACMID-Cleaned dataset. Leverag- ing abundant data, we expand the conventional classifica- tion from 4-stem (Vocal/Bass/Drums/Others) to 7-stem (Pi- ano/Drums/Bass/Acoustic Guitar/Electric Guitar/Strings/Wind- Brass), enabling high granularity MSS systems. Experiments on SOTA MSS model demonstrates two key results: (i) MSS model trained with ACMID-Cleaned achieved a 2.39dB improvement in SDR performance compared to that with ACMID-Uncleaned, demostrating the effectiveness of our data cleaning procedure; (ii) incorporating ACMID-Cleaned to training enhances MSS model's average performance by 1.16dB, confirming the value of our dataset. Our data crawl- ing code, cleaning model code and weights are available at: https://github.com/scottishfold0621/ACMID.

☆ Full-Duplex-Bench-v2: A Multi-Turn Evaluation Framework for Duplex Dialogue Systems with an Automated Examiner

Guan-Ting Lin, Shih-Yun Shan Kuan, Jiatong Shi, Kai-Wei Chang, Siddhant Arora, Shinji Watanabe, Hung-yi Lee

While full-duplex speech agents enable natural, low-latency interaction by speaking and listening simultaneously, their consistency and task performance in multi-turn settings remain underexplored. We introduce Full-Duplex-Bench-v2 (FDB-v2), a streaming framework that integrates with an automated examiner that enforces staged goals under two pacing setups (Fast vs. Slow). FDB-v2 covers four task families: daily, correction, entity tracking, and safety. We report turn-taking fluency, multi-turn instruction following, and task-specific competence. The framework is extensible, supporting both commercial APIs and open source models. When we test full-duplex systems with FDB-v2, they often get confused when people talk at the same time, struggle to handle corrections smoothly, and sometimes lose track of who or what is being talked about. Through an open-sourced, standardized streaming protocol and a task set, FDB-v2 makes it easy to extend to new task families, allowing the community to tailor and accelerate evaluation of multi-turn full-duplex systems.

comment: Work in progress

☆ ACMID: Automatic Curation of Musical Instrument Dataset for 7-Stem Music Source Separation

Ji Yu, Yang shuo, Xu Yuetonghui, Liu Mengmei, Ji Qiang, Han Zerui

Most current music source separation (MSS) methods rely on supervised learning, limited by training data quantity and quality. Though web-crawling can bring abundant data, platform-level track labeling often causes metadata mismatches, impeding accurate "audio-label" pair acquisition. To address this, we present ACMID: a dataset for MSS generated through web crawling of extensive raw data, followed by automatic cleaning via an instrument classifier built on a pre-trained audio encoder that filters and aggregates clean segments of target instruments from the crawled tracks, resulting in the refined ACMID-Cleaned dataset. Leveraging abundant data, we expand the conventional classification from 4-stem (Vocal/Bass/Drums/Others) to 7-stem (Piano/Drums/Bass/Acoustic Guitar/Electric Guitar/Strings/Wind-Brass), enabling high granularity MSS systems. Experiments on SOTA MSS model demonstrates two key results: (i) MSS model trained with ACMID-Cleaned achieved a 2.39dB improvement in SDR performance compared to that with ACMID-Uncleaned, demostrating the effectiveness of our data cleaning procedure; (ii) incorporating ACMID-Cleaned to training enhances MSS model's average performance by 1.16dB, confirming the value of our dataset. Our data crawling code, cleaning model code and weights are available at: https://github.com/scottishfold0621/ACMID.

☆ Audible Networks: Deconstructing and Manipulating Sounds with Deep Non-Negative Autoencoders

Juan José Burred, Carmine-Emanuele Cella

We propose the use of Non-Negative Autoencoders (NAEs) for sound deconstruction and user-guided manipulation of sounds for creative purposes. NAEs offer a versatile and scalable extension of traditional Non-Negative Matrix Factorization (NMF)-based approaches for interpretable audio decomposition. By enforcing non-negativity constraints through projected gradient descent, we obtain decompositions where internal weights and activations can be directly interpreted as spectral shapes and temporal envelopes, and where components can themselves be listened to as individual sound events. In particular, multi-layer Deep NAE architectures enable hierarchical representations with an adjustable level of granularity, allowing sounds to be deconstructed at multiple levels of abstraction: from high-level note envelopes down to fine-grained spectral details. This framework enables a wide new range of expressive, controllable, and randomized sound transformations. We introduce novel manipulation operations including cross-component and cross-layer synthesis, hierarchical deconstructions, and several randomization strategies that control timbre and event density. Through visualizations and resynthesis of practical examples, we demonstrate how NAEs can serve as flexible and interpretable tools for object-based sound editing.

☆ Beyond Discrete Categories: Multi-Task Valence-Arousal Modeling for Pet Vocalization Analysis

Junyao Huang, Rumin Situ

Traditional pet emotion recognition from vocalizations, based on discrete classification, struggles with ambiguity and capturing intensity variations. We propose a continuous Valence-Arousal (VA) model that represents emotions in a two-dimensional space. Our method uses an automatic VA label generation algorithm, enabling large-scale annotation of 42,553 pet vocalization samples. A multi-task learning framework jointly trains VA regression with auxiliary tasks (emotion, body size, gender) to enhance prediction by improving feature learning. Our Audio Transformer model achieves a validation Valence Pearson correlation of r = 0.9024 and an Arousal r = 0.7155, effectively resolving confusion between discrete categories like "territorial" and "happy." This work introduces the first continuous VA framework for pet vocalization analysis, offering a more expressive representation for human-pet interaction, veterinary diagnostics, and behavioral training. The approach shows strong potential for deployment in consumer products like AI pet emotion translators.

comment: 24 pages, 6 figures, 4 tables. First continuous VA framework for pet vocalization analysis with 42,553 samples

♻ ☆ I$^2$RF-TFCKD: Intra-Inter Representation Fusion with Time-Frequency Calibration Knowledge Distillation for Speech Enhancement

Jiaming Cheng, Ruiyu Liang, Ye Ni, Chao Xu, Jing Li, Wei Zhou, Rui Liu, Björn W. Schuller, Xiaoshuai Hao

In this paper, we propose an intra-inter representation fusion knowledge distillation (KD) framework with time-frequency calibration (I$^2$RF-TFCKD) for SE, which achieves distillation through the fusion of multi-layer teacher-student feature flows. Different from previous distillation strategies for SE, the proposed framework fully utilizes the time-frequency differential information of speech while promoting global knowledge flow. Firstly, we construct a collaborative distillation paradigm for intra-set and inter-set correlations. Within a correlated set, multi-layer teacher-student features are pairwise matched for calibrated distillation. Subsequently, we generate representative features from each correlated set through residual fusion to form the fused feature set that enables inter-set knowledge interaction. Secondly, we propose a multi-layer interactive distillation based on dual-stream time-frequency cross-calibration, which calculates the teacher-student similarity calibration weights in the time and frequency domains respectively and performs cross-weighting, thus enabling refined allocation of distillation contributions across different layers according to speech characteristics. The proposed distillation strategy is applied to the dual-path dilated convolutional recurrent network (DPDCRN) that ranked first in the SE track of the L3DAS23 challenge. To evaluate the effectiveness of I$^2$RF-TFCKD, we conduct experiments on both single-channel and multi-channel SE datasets. Objective evaluations demonstrate that the proposed KD strategy consistently and effectively improves the performance of the low-complexity student model and outperforms other distillation schemes.

comment: submitted to Information Fusion

♻ ☆ Towards Frame-level Quality Predictions of Synthetic Speech

Michael Kuhlmann, Fritz Seebauer, Petra Wagner, Reinhold Haeb-Umbach

While automatic subjective speech quality assessment has witnessed much progress, an open question is whether an automatic quality assessment at frame resolution is possible. This would be highly desirable, as it adds explainability to the assessment of speech synthesis systems. Here, we take first steps towards this goal by identifying issues of existing quality predictors that prevent sensible frame-level prediction. Further, we define criteria that a frame-level predictor should fulfill. We also suggest a chunk-based processing that avoids the impact of a localized distortion on the score of neighboring frames. Finally, we measure in experiments with localized artificial distortions the localization performance of a set of frame-level quality predictors and show that they can outperform detection performance of human annotations obtained from a crowd-sourced perception experiment.

comment: Proceedings of Interspeech

♻ ☆ STOPA: A Database of Systematic VariaTion Of DeePfake Audio for Open-Set Source Tracing and Attribution

Anton Firc, Manasi Chhibber, Jagabandhu Mishra, Vishwanath Pratap Singh, Tomi Kinnunen, Kamil Malinka

A key research area in deepfake speech detection is source tracing - determining the origin of synthesised utterances. The approaches may involve identifying the acoustic model (AM), vocoder model (VM), or other generation-specific parameters. However, progress is limited by the lack of a dedicated, systematically curated dataset. To address this, we introduce STOPA, a systematically varied and metadata-rich dataset for deepfake speech source tracing, covering 8 AMs, 6 VMs, and diverse parameter settings across 700k samples from 13 distinct synthesisers. Unlike existing datasets, which often feature limited variation or sparse metadata, STOPA provides a systematically controlled framework covering a broader range of generative factors, such as the choice of the vocoder model, acoustic model, or pretrained weights, ensuring higher attribution reliability. This control improves attribution accuracy, aiding forensic analysis, deepfake detection, and generative model transparency.

comment: Published at Interspeech 2025 conference

♻ ☆ Audio-Conditioned Diffusion LLMs for ASR and Deliberation Processing

Mengqi Wang, Zhan Liu, Zengrui Jin, Guangzhi Sun, Chao Zhang, Philip C. Woodland

Diffusion-based large language models (DLLMs) have recently attracted growing interest as an alternative to autoregressive decoders. In this work, we present an empirical study on using the diffusion-based large language model LLaDA for automatic speech recognition (ASR). We first investigate its use as an external deliberation-based processing module for Whisper-LLaMA transcripts. By leveraging the bidirectional attention and denoising capabilities of LLaDA, we explore random masking, low-confidence masking, and semi-autoregressive strategies, showing that Whisper-LLaDA substantially reduces WER compared with the baseline. On LibriSpeech, the best cascade system achieves 2.25%/4.94% WER on test-clean/test-other, representing a 12.3% relative improvement over the Whisper-LLaMA baseline on the test-other split. In contrast, a plain-text LLaDA without acoustic features fails to improve accuracy, highlighting the importance of audio-conditioned embeddings. We further evaluate Whisper-LLaDA as a standalone decoder for ASR with diffusion-based and semi-autoregressive decoding. Most experimental configurations achieve faster inference than the Whisper-LLaMA baseline, although recognition accuracy is slightly lower. These findings offer an empirical view of diffusion-based LLMs for ASR and point to promising directions for improvements.

♻ ☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026

Vaibhav Srivastav, Steven Zheng, Eric Bezzam, Eustache Le Bihan, Nithin Koluguri, Piotr Żelasko, Somshubra Majumdar, Adel Moumen, Sanchit Gandhi

Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard ; Code: https://github.com/huggingface/open_asr_leaderboard

♻ ☆ SeamlessEdit: Background Noise Aware Zero-Shot Speech Editing with in-Context Enhancement

Kuan-Yu Chen, Jeng-Lin Li, Jian-Jiun Ding

With the fast development of zero-shot text-to-speech technologies, it is possible to generate high-quality speech signals that are indistinguishable from the real ones. Speech editing, including speech insertion and replacement, appeals to researchers due to its potential applications. However, existing studies only considered clean speech scenarios. In real-world applications, the existence of environmental noise could significantly degrade the quality of generation. In this study, we propose a noise-resilient speech editing framework, SeamlessEdit, for noisy speech editing. SeamlessEdit adopts a frequency-band-aware noise suppression module and an in-content refinement strategy. It can well address the scenario where the frequency bands of voice and background noise are not separated. The proposed SeamlessEdit framework outperforms state-of-the-art approaches in multiple quantitative and qualitative evaluations.

comment: 5 pages, 3 figures

☆ Comparison of Speech Tasks in Human Expert and Machine Detection of Parkinson's Disease

Peter Plantinga, Roozbeh Sattari, Karine Marcotte, Carla Di Gironimo, Madeleine Sharp, Liziane Bouvier, Maiya Geddes, Ingrid Verduyckt, Étienne de Villers-Sidani, Mirco Ravanelli, Denise Klein

The speech of people with Parkinson's Disease (PD) has been shown to hold important clues about the presence and progression of the disease. We investigate the factors based on which humans experts make judgments of the presence of disease in speech samples over five different speech tasks: phonations, sentence repetition, reading, recall, and picture description. We make comparisons by conducting listening tests to determine clinicians accuracy at recognizing signs of PD from audio alone, and we conduct experiments with a machine learning system for detection based on Whisper. Across tasks, Whisper performs on par or better than human experts when only audio is available, especially on challenging but important subgroups of the data: younger patients, mild cases, and female patients. Whisper's ability to recognize acoustic cues in difficult cases complements the multimodal and contextual strengths of human experts.

comment: Accepted to SMASH 2025

☆ AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs

Peize He, Zichen Wen, Yubo Wang, Yuxuan Wang, Xiaoqian Liu, Jiajie Huang, Zehui Lei, Zhuangcheng Gu, Xiangqi Jin, Jiabing Yang, Kai Li, Zhifei Liu, Weijia Li, Cunxiang Wang, Conghui He, Linfeng Zhang

Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.

comment: 26 pages, 23 figures, the code is available at \url{https://github.com/DabDans/AudioMarathon}

☆ Making Machines Sound Sarcastic: LLM-Enhanced and Retrieval-Guided Sarcastic Speech Synthesis

Zhu Li, Yuqing Zhang, Xiyuan Gao, Shekhar Nayak, Matt Coler

Sarcasm is a subtle form of non-literal language that poses significant challenges for speech synthesis due to its reliance on nuanced semantic, contextual, and prosodic cues. While existing speech synthesis research has focused primarily on broad emotional categories, sarcasm remains largely unexplored. In this paper, we propose a Large Language Model (LLM)-enhanced Retrieval-Augmented framework for sarcasm-aware speech synthesis. Our approach combines (1) semantic embeddings from a LoRA-fine-tuned LLaMA 3, which capture pragmatic incongruity and discourse-level cues of sarcasm, and (2) prosodic exemplars retrieved via a Retrieval Augmented Generation (RAG) module, which provide expressive reference patterns of sarcastic delivery. Integrated within a VITS backbone, this dual conditioning enables more natural and contextually appropriate sarcastic speech. Experiments demonstrate that our method outperforms baselines in both objective measures and subjective evaluations, yielding improvements in speech naturalness, sarcastic expressivity, and downstream sarcasm detection.

☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026

Vaibhav Srivastav, Steven Zheng, Eric Bezzam, Eustache Le Bihan, Nithin Koluguri, Piotr Żelasko, Somshubra Majumdar, Adel Moumen, Sanchit Gandhi

Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard; Code: https://github.com/huggingface/open_asr_leaderboard

☆ Moises-Light: Resource-efficient Band-split U-Net For Music Source Separation

Yun-Ning, Hung, Igor Pereira, Filip Korzeniowski

In recent years, significant advances have been made in music source separation, with model architectures such as dual-path modeling, band-split modules, or transformer layers achieving comparably good results. However, these models often contain a significant number of parameters, posing challenges to devices with limited computational resources in terms of training and practical application. While some lightweight models have been introduced, they generally perform worse compared to their larger counterparts. In this paper, we take inspiration from these recent advances to improve a lightweight model. We demonstrate that with careful design, a lightweight model can achieve comparable SDRs to models with up to 13 times more parameters. Our proposed model, Moises-Light, achieves competitive results in separating four musical stems on the MUSDB-HQ benchmark dataset. The proposed model also demonstrates competitive scalability when using MoisesDB as additional training data.

☆ XLSR-Kanformer: A KAN-Intergrated model for Synthetic Speech Detection

Phuong Tuan Dat, Tran Huy Dat

Recent advancements in speech synthesis technologies have led to increasingly sophisticated spoofing attacks, posing significant challenges for automatic speaker verification systems. While systems based on self-supervised learning (SSL) models, particularly the XLSR-Conformer architecture, have demonstrated remarkable performance in synthetic speech detection, there remains room for architectural improvements. In this paper, we propose a novel approach that replaces the traditional Multi-Layer Perceptron (MLP) in the XLSR-Conformer model with a Kolmogorov-Arnold Network (KAN), a powerful universal approximator based on the Kolmogorov-Arnold representation theorem. Our experimental results on ASVspoof2021 demonstrate that the integration of KAN to XLSR-Conformer model can improve the performance by 60.55% relatively in Equal Error Rate (EER) LA and DF sets, further achieving 0.70% EER on the 21LA set. Besides, the proposed replacement is also robust to various SSL architectures. These findings suggest that incorporating KAN into SSL-based models is a promising direction for advances in synthetic speech detection.

comment: Accepted to 2025 IEEE International Conference on Advanced Video and Signal-Based Surveillance

☆ Pitch Estimation With Mean Averaging Smoothed Product Spectrum And Musical Consonance Evaluation Using MASP

Murat Yasar Baskin

This study introduces Mean Averaging Smoothed Product (MASP) Spectrum, which is a modified version of the Harmonic Product Spectrum, designed to enhance pitch estimation for many algorithm-wise deceptive frequency spectra that still lead clear pitches, for both harmonic and inharmonic cases. By introducing a global mean based smoothing for spectrum, the MASP algorithm diminishes the unwanted sensitivity of HPS for spectra with missing partials. The method exhibited robust pitch estimations consistent with perceptual expectations. Motivated upon the strong correlation between consonance and periodicity, the same algorithm is extended and, with the proposition of a harmonicity measure (H), used to evaluate musical consonance for two and three tones; yielding consonance hierarchies that align with perception and practice of music theory. These findings suggest that perception of pitch and consonance may share a similar underlying mechanism that depend on spectrum.

☆ Benchmarking Fake Voice Detection in the Fake Voice Generation Arms Race

Xutao Mao, Ke Li, Cameron Baird, Ezra Xuanru Tao, Dan Lin

As advances in synthetic voice generation accelerate, an increasing variety of fake voice generators have emerged, producing audio that is often indistinguishable from real human speech. This evolution poses new and serious threats across sectors where audio recordings serve as critical evidence. Although fake voice detectors are also advancing, the arms race between fake voice generation and detection has become more intense and complex. In this work, we present the first large-scale, cross-domain evaluation of fake voice detectors, benchmarking 8 state-of-the-art models against datasets synthesized by 20 different fake voice generation systems. To the best of our knowledge, this is the most comprehensive cross-domain assessment conducted to date. Our study reveals substantial security vulnerabilities in current fake voice detection systems, underscoring critical gaps in their real-world robustness. To advance the field, we propose a unified and effective metric that consolidates the diverse and often inconsistent evaluation criteria previously used across different studies. This metric enables standardized, straightforward comparisons of the robustness of fake voice detectors. We conclude by offering actionable recommendations for building more resilient fake voice detection technologies, with the broader goal of reinforcing the foundations of AI security and trustworthiness.

☆ BACHI: Boundary-Aware Symbolic Chord Recognition Through Masked Iterative Decoding on Pop and Classical Music

Mingyang Yao, Ke Chen, Shlomo Dubnov, Taylor Berg-Kirkpatrick

Automatic chord recognition (ACR) via deep learning models has gradually achieved promising recognition accuracy, yet two key challenges remain. First, prior work has primarily focused on audio-domain ACR, while symbolic music (e.g., score) ACR has received limited attention due to data scarcity. Second, existing methods still overlook strategies that are aligned with human music analytical practices. To address these challenges, we make two contributions: (1) we introduce POP909-CL, an enhanced version of POP909 dataset with tempo-aligned content and human-corrected labels of chords, beats, keys, and time signatures; and (2) We propose BACHI, a symbolic chord recognition model that decomposes the task into different decision steps, namely boundary detection and iterative ranking of chord root, quality, and bass (inversion). This mechanism mirrors the human ear-training practices. Experiments demonstrate that BACHI achieves state-of-the-art chord recognition performance on both classical and pop music benchmarks, with ablation studies validating the effectiveness of each module.

comment: Under review

☆ INFER : Learning Implicit Neural Frequency Response Fields for Confined Car Cabin

Harshvardhan C. Takawale, Nirupam Roy, Phil Brown

Accurate modeling of spatial acoustics is critical for immersive and intelligible audio in confined, resonant environments such as car cabins. Current tuning methods are manual, hardware-intensive, and static, failing to account for frequency selective behaviors and dynamic changes like passenger presence or seat adjustments. To address this issue, we propose INFER: Implicit Neural Frequency Response fields, a frequency-domain neural framework that is jointly conditioned on source and receiver positions, orientations to directly learn complex-valued frequency response fields inside confined, resonant environments like car cabins. We introduce three key innovations over current neural acoustic modeling methods: (1) novel end-to-end frequency-domain forward model that directly learns the frequency response field and frequency-specific attenuation in 3D space; (2) perceptual and hardware-aware spectral supervision that emphasizes critical auditory frequency bands and deemphasizes unstable crossover regions; and (3) a physics-based Kramers-Kronig consistency constraint that regularizes frequency-dependent attenuation and delay. We evaluate our method over real-world data collected in multiple car cabins. Our approach significantly outperforms time- and hybrid-domain baselines on both simulated and real-world automotive datasets, cutting average magnitude and phase reconstruction errors by over 39% and 51%, respectively. INFER sets a new state-of-the-art for neural acoustic modeling in automotive spaces

☆ AV-EMO-Reasoning: Benchmarking Emotional Reasoning Capabilities in Omni-modal LLMS with Audio-visual Cues

Krish Patel, Dingkun Zhou, Ajay Kankipati, Akshaj Gupta, Zeyi Austin Li, Mohul Shukla, Vibhor Narang, Sara Kofman, Zongli Ye, Grace Wang, Xiaoyu Shi, Tingle Li, Guan-Ting Lin, Kan Jen Cheng, Huang-Cheng Chou, Jiachen Lian, Gopala Anumanchipalli

Emotions conveyed through voice and face shape engagement and context in human-AI interaction. Despite rapid progress in omni-modal large language models (LLMs), the holistic evaluation of emotional reasoning with audiovisual cues remains limited. To address this gap, we introduce AV-EMO-Reasoning, a benchmark designed to systematically assess emotional coherence in LLMs. The framework leverages a curated, single- and multi-turn synthetic audiovisual corpus with a real-world set and is assessed under continuous, categorical, and perceptual metrics. Experiments with leading LLMs show that visual cues reliably improve emotional coherence over audio-only baselines. Moreover, LLMs can leverage audio-visual cues to generate more emotion-aware speech. Models exhibit complementary strengths across metric families, indicating that automatic scores capture facets distinct from perceptual judgments. By releasing a systematic evaluation benchmark, AV-EMO-Reasoning offers a reproducible standard for evaluating emotion-aware dialogue and advances toward more natural, adaptive human-AI interaction.

☆ Look before Transcription: End-to-End SlideASR with Visually-Anchored Policy Optimization

Rui Hu, Delai Qiu, Yining Wang, Shengping Liu, Jitao Sang

Automatic speech recognition (ASR) systems often struggle with domain-specific terminology, especially in specialized settings such as academic lectures. To address this, we define the SlideASR task, which leverages the rich visual information from presentation slides to improve transcription accuracy. Existing pipeline methods for this task tend to be complex and underperform. Although omni-modal large language models (OLLMs) provide a promising end-to-end framework, they frequently fail in practice by degenerating into simple optical character recognition (OCR) systems. To overcome this, we propose Visually-Anchored Policy Optimization (VAPO), a novel post-training method designed to control the model's reasoning process. Drawing on the Chain-of-Thought reasoning paradigm, VAPO enforces a structured "Look before Transcription" procedure using a format. Specifically, the model first performs OCR on the slide content within the think step, then generates the transcription by referencing this recognized visual information in the answer step. This reasoning process is optimized via reinforcement learning with four distinct rewards targeting format compliance, OCR accuracy, ASR quality, and visual anchoring consistency. To support further research, we construct SlideASR-Bench, a new entity-rich benchmark consisting of a synthetic dataset for training and testing, and a challenging real-world set for evaluation. Extensive experiments demonstrate that VAPO significantly improves recognition of domain-specific terms, establishing an effective end-to-end paradigm for SlideASR.

♻ ☆ LaunchpadGPT: Language Model as Music Visualization Designer on Launchpad

Siting Xu, Yolo Yunlong Tang, Feng Zheng

Launchpad is a musical instrument that allows users to create and perform music by pressing illuminated buttons. To assist and inspire the design of the Launchpad light effect, and provide a more accessible approach for beginners to create music visualization with this instrument, we proposed the LaunchpadGPT model to generate music visualization designs on Launchpad automatically. Based on the language model with excellent generation ability, our proposed LaunchpadGPT takes an audio piece of music as input and outputs the lighting effects of Launchpad-playing in the form of a video (Launchpad-playing video). We collect Launchpad-playing videos and process them to obtain music and corresponding video frame of Launchpad-playing as prompt-completion pairs, to train the language model. The experiment result shows the proposed method can create better music visualization than random generation methods and hold the potential for a broader range of music visualization applications. Our code is available at https://github.com/yunlong10/LaunchpadGPT/.

comment: Accepted to International Computer Music Conference (ICMC) 2023

♻ ☆ AbsoluteNet: A Deep Learning Neural Network to Classify Cerebral Hemodynamic Responses of Auditory Processing

Behtom Adeli, John Mclinden, Pankaj Pandey, Ming Shao, Yalda Shahriari

In recent years, deep learning (DL) approaches have demonstrated promising results in decoding hemodynamic responses captured by functional near-infrared spectroscopy (fNIRS), particularly in the context of brain-computer interface (BCI) applications. This work introduces AbsoluteNet, a novel deep learning architecture designed to classify auditory event-related responses recorded using fNIRS. The proposed network is built upon principles of spatio-temporal convolution and customized activation functions. Our model was compared against several models, namely fNIRSNET, MDNN, DeepConvNet, and ShallowConvNet. The results showed that AbsoluteNet outperforms existing models, reaching 87.0% accuracy, 84.8% sensitivity, and 89.2% specificity in binary classification, surpassing fNIRSNET, the second-best model, by 3.8% in accuracy. These findings underscore the effectiveness of our proposed deep learning model in decoding hemodynamic responses related to auditory processing and highlight the importance of spatio-temporal feature aggregation and customized activation functions to better fit fNIRS dynamics.

♻ ☆ The Sound of Syntax: Finetuning and Comprehensive Evaluation of Language Models for Speech Pathology EMNLP 2025

Fagun Patel, Duc Q. Nguyen, Sang T. Truong, Jody Vaynshtok, Sanmi Koyejo, Nick Haber

According to the U.S. National Institutes of Health, more than 3.4 million children experience speech disorders that require clinical intervention. The number of speech-language pathologists (SLPs) is roughly 20 times fewer than the number of affected children, highlighting a significant gap in children's care and a pressing need for technological support that improves the productivity of SLPs. State-of-the-art multimodal language models (MLMs) show promise for supporting SLPs, but their use remains underexplored largely due to a limited understanding of their performance in high-stakes clinical settings. To address this gap, we collaborate with domain experts to develop a taxonomy of real-world use cases of MLMs in speech-language pathologies. Building on this taxonomy, we introduce the first comprehensive benchmark for evaluating MLM across five core use cases, each containing 1,000 manually annotated data points. This benchmark includes robustness and sensitivity tests under various settings, including background noise, speaker gender, and accent. Our evaluation of 15 state-of-the-art MLMs reveals that no single model consistently outperforms others across all tasks. Notably, we find systematic disparities, with models performing better on male speakers, and observe that chain-of-thought prompting can degrade performance on classification tasks with large label spaces and narrow decision boundaries. Furthermore, we study fine-tuning MLMs on domain-specific data, achieving improvements of over 10\% compared to base models. These findings highlight both the potential and limitations of current MLMs for speech-language pathology applications, underscoring the need for further research and targeted development.

comment: EMNLP 2025 Oral Presentation

♻ ☆ A Differentiable Alignment Framework for Sequence-to-Sequence Modeling via Optimal Transport

Yacouba Kaloga, Shashi Kumar, Petr Motlicek, Ina Kodrasi

Accurate sequence-to-sequence (seq2seq) alignment is critical for applications like medical speech analysis and language learning tools relying on automatic speech recognition (ASR). State-of-the-art end-to-end (E2E) ASR systems, such as the Connectionist Temporal Classification (CTC) and transducer-based models, suffer from peaky behavior and alignment inaccuracies. In this paper, we propose a novel differentiable alignment framework based on one-dimensional optimal transport, enabling the model to learn a single alignment and perform ASR in an E2E manner. We introduce a pseudo-metric, called Sequence Optimal Transport Distance (SOTD), over the sequence space and discuss its theoretical properties. Based on the SOTD, we propose Optimal Temporal Transport Classification (OTTC) loss for ASR and contrast its behavior with CTC. Experimental results on the TIMIT, AMI, and LibriSpeech datasets show that our method considerably improves alignment performance compared to CTC and the more recently proposed Consistency-Regularized CTC, though with a trade-off in ASR performance. We believe this work opens new avenues for seq2seq alignment research, providing a solid foundation for further exploration and development within the community.

♻ ☆ Enhancing Few-shot Keyword Spotting Performance through Pre-Trained Self-supervised Speech Models

Alican Gok, Oguzhan Buyuksolak, Osman Erman Okman, Murat Saraclar

Keyword Spotting plays a critical role in enabling hands-free interaction for battery-powered edge devices. Few-Shot Keyword Spotting (FS-KWS) addresses the scalability and adaptability challenges of traditional systems by enabling recognition of custom keywords with only a few examples. However, existing FS-KWS systems achieve subpar accuracy at desirable false acceptance rates, particularly in resource-constrained edge environments. To address these issues, we propose a training scheme that leverages self-supervised learning models for robust feature extraction, dimensionality reduction, and knowledge distillation. The teacher model, based on Wav2Vec 2.0 is trained using Sub-center ArcFace loss, which enhances inter-class separability and intra-class compactness. To enable efficient deployment on edge devices, we introduce attention-based dimensionality reduction and train a standard lightweight ResNet15 student model. We evaluate the proposed approach on the English portion of the Multilingual Spoken Words Corpus (MSWC) and the Google Speech Commands (GSC) datasets. Notably, the proposed training method improves the 10-shot classification accuracy from 33.4% to 74.1% on 11 classes at 1% false alarm accuracy on the GSC dataset, thus making it significantly better-suited for a real use case scenario.

comment: Submitted to IEEE Signal Processing Letters, 5 pages, 3 figures

♻ ☆ Token-based Audio Inpainting via Discrete Diffusion

Tali Dror, Iftach Shoham, Moshe Buchris, Oren Gal, Haim Permuter, Gilad Katz, Eliya Nachmani

Audio inpainting seeks to restore missing segments in degraded recordings. Previous diffusion-based methods exhibit impaired performance when the missing region is large. We introduce the first approach that applies discrete diffusion over tokenized music representations from a pre-trained audio tokenizer, enabling stable and semantically coherent restoration of long gaps. Our method further incorporates two training approaches: a derivative-based regularization loss that enforces smooth temporal dynamics, and a span-based absorbing transition that provides structured corruption during diffusion. Experiments on the MusicNet and MAESTRO datasets with gaps up to 750 ms show that our approach consistently outperforms strong baselines across range of gap lengths, for gaps of 150 ms and above. This work advances musical audio restoration and introduces new directions for discrete diffusion model training. Audio examples of our proposed method can be found at https://iftach21.github.io/.

♻ ☆ Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation

Haorui He, Zengqiang Shang, Chaoren Wang, Xuyuan Li, Yicheng Gu, Hua Hua, Liwei Liu, Chen Yang, Jiaqi Li, Peiyang Shi, Yuancheng Wang, Kai Chen, Pengyuan Zhang, Zhizheng Wu

Recent advancements in speech generation have been driven by large-scale training datasets. However, current models struggle to capture the spontaneity and variability inherent in real-world human speech, as they are primarily trained on audio-book datasets limited to formal, read-aloud speaking styles. To address this limitation, we introduce Emilia-Pipe, an open-source preprocessing pipeline designed to extract high-quality training data from valuable yet under-explored in-the-wild sources that capture spontaneous human speech in real-world contexts. Using Emilia-Pipe, we construct Emilia, which comprises over 101k hours of speech across six languages: English, Chinese, German, French, Japanese, and Korean. Furthermore, we expand Emilia to Emilia-Large, a dataset exceeding 216k hours, making it one of the largest open-source speech generation resources available. Extensive experiments show that Emilia-trained models produce markedly more spontaneous, human-like speech than those trained on traditional audio-book datasets, while matching their intelligibility. These models better capture diverse speaker timbres and the full spectrum of real-world conversational styles. Our work also highlights the importance of scaling dataset size for advancing speech generation performance and validates the effectiveness of Emilia for both multilingual and crosslingual speech generation tasks.

comment: Full version of arXiv:2407.05361, dataset is available at: https://huggingface.co/datasets/amphion/Emilia-Dataset

♻ ☆ PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction

Shufan Li, Aditya Grover

Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses. However, their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user experiences. This latency is particularly evident when LLMs are deployed as single-user voice assistants on consumer-grade hardware with limited computing capacity. We discovered that this latency is primarily dominated by the time it takes for the LLMs to generate the first sentence, which is required as input by the TTS systems that synthesize audio responses on a sentence-by-sentence basis. To address this bottleneck, we propose Predictive Generation (PredGen), a novel framework that mitigates-or even eliminates-this delay through speculative decoding at input time. PredGen generates candidate responses while the user is still speaking, enabling the system to begin TTS processing with minimal delay. Simulated experiments on the Lmsys and MT-Bench datasets show that the proposed method can effectively reduce the latency by around 2x across a wide range of use cases, while incurring only minimal additional computation cost at input time-computation that would otherwise go unused.

comment: 16 pages,4 figures

♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS

Seungheon Doh, Keunwoo Choi, Juhan Nam

While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.

comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)

♻ ☆ Self-Supervised Speech Quality Assessment (S3QA): Leveraging Speech Foundation Models for a Scalable Speech Quality Metric

Mattson Ogg, Caitlyn Bishop, Han Yi, Sarah Robinson

Methods for automatically assessing speech quality in real world environments are critical for developing robust human language technologies and assistive devices. Behavioral ratings provided by human raters (e.g., mean opinion scores; MOS) are considered the gold standard, but they are susceptible to variability between individual raters, cannot easily be generalized across corpora, and are labor-intensive to collect, thus limiting the acoustic challenges they can quantify. Here, we present a new, scalable method for automatically assessing speech quality: the self-supervised speech quality assessment (S3QA) model. First, we manipulated high quality utterances from multiple speech corpora, using a wide range of acoustic challenges intended to emulate common sources of quality degradation in the real-world: frequency filtering, reverberation, background noise, and digital compression. Second, we leveraged an existing, pre-trained speech foundation model, WavLM, to computationally derive a self-supervised training target that quantified speech degradation using the cosine distance between the clean and degraded versions of each utterance in the embedding space. Next, we trained a transformer-based model to predict these cosine distances, given only the degraded versions of the utterances. Finally, the trained model was evaluated on unseen test corpora of synthetic mixtures, NISQA, and VOiCES. We show that the S3QA model trained on this task accurately predicts degradation cosine distances across a wide range challenging acoustic conditions and is aligned with both behavioral ratings (MOS), speech technology performance (automatic speech recognition) and other important features of the held-out data (e.g., microphone distances). This model provides an automated, scalable method for assessing speech quality across a wide range of acoustic challenges.

comment: 1 table, seven figures, thirteen pages

♻ ☆ TalkPlayData 2: An Agentic Synthetic Data Pipeline for Multimodal Conversational Music Recommendation

Keunwoo Choi, Seungheon Doh, Juhan Nam

We present TalkPlayData 2, a synthetic dataset for multimodal conversational music recommendation generated by an agentic data pipeline. In the proposed pipeline, multiple large language model (LLM) agents are created under various roles with specialized prompts and access to different parts of information, and the chat data is acquired by logging the conversation between the Listener LLM and the Recsys LLM. To cover various conversation scenarios, for each conversation, the Listener LLM is conditioned on a finetuned conversation goal. Finally, all the LLMs are multimodal with audio and images, allowing a simulation of multimodal recommendation and conversation. In the LLM-as-a-judge and subjective evaluation experiments, TalkPlayData 2 achieved the proposed goal in various aspects related to training a generative recommendation model for music. TalkPlayData 2 and its generation code are released at https://talkpl.ai/talkplaydata2.

comment: 2025-10-08: updating the stat table with the latest numbers. updated the abstract per the latest license terms

♻ ☆ Evaluating Sound Similarity Metrics for Differentiable, Iterative Sound-Matching

Amir Salimi, Abram Hindle, Osmar R. Zaiane

Manual sound design with a synthesizer is inherently iterative: an artist compares the synthesized output to a mental target, adjusts parameters, and repeats until satisfied. Iterative sound-matching automates this workflow by continually programming a synthesizer under the guidance of a loss function (or similarity measure) toward a target sound. Prior comparisons of loss functions have typically favored one metric over another, but only within narrow settings: limited synthesis methods, few loss types, often without blind listening tests. This leaves open the question of whether a universally optimal loss exists, or the choice of loss remains a creative decision conditioned on the synthesis method and the sound designer's preference. We propose differentiable iterative sound-matching as the natural extension of the available literature, since it combines the manual approach to sound design with modern advances in machine learning. To analyze the variability of loss function performance across synthesizers, we implemented a mix of four novel and established differentiable loss functions, and paired them with differentiable subtractive, additive, and AM synthesizers. For each of the sixteen synthesizer--loss combinations, we ran 300 randomized sound-matching trials. Performance was measured using parameter differences, spectrogram-distance metrics, and manually assigned listening scores. We observed a moderate level of consistency among the three performance measures. Our post-hoc analysis shows that the loss function performance is highly dependent on the synthesizer. These findings underscore the value of expanding the scope of sound-matching experiments and developing new similarity metrics tailored to specific synthesis techniques rather than pursuing one-size-fits-all solutions.

♻ ☆ Multi-Target Backdoor Attacks Against Speaker Recognition

Alexandrine Fortier, Sonal Joshi, Thomas Thebaud, Jesús Villalba, Najim Dehak, Patrick Cardinal

In this work, we propose a multi-target backdoor attack against speaker identification using position-independent clicking sounds as triggers. Unlike previous single-target approaches, our method targets up to 50 speakers simultaneously, achieving success rates of up to 95.04%. To simulate more realistic attack conditions, we vary the signal-to-noise ratio between speech and trigger, demonstrating a trade-off between stealth and effectiveness. We further extend the attack to the speaker verification task by selecting the most similar training speaker - based on cosine similarity - as a proxy target. The attack is most effective when target and enrolled speaker pairs are highly similar, reaching success rates of up to 90% in such cases.

comment: Accepted to IEEE Automatic Speech Recognition and Understanding Workshop 2025

☆ Comparison of Speech Tasks in Human Expert and Machine Detection of Parkinson's Disease

Peter Plantinga, Roozbeh Sattari, Karine Marcotte, Carla Di Gironimo, Madeleine Sharp, Liziane Bouvier, Maiya Geddes, Ingrid Verduyckt, Étienne de Villers-Sidani, Mirco Ravanelli, Denise Klein

The speech of people with Parkinson's Disease (PD) has been shown to hold important clues about the presence and progression of the disease. We investigate the factors based on which humans experts make judgments of the presence of disease in speech samples over five different speech tasks: phonations, sentence repetition, reading, recall, and picture description. We make comparisons by conducting listening tests to determine clinicians accuracy at recognizing signs of PD from audio alone, and we conduct experiments with a machine learning system for detection based on Whisper. Across tasks, Whisper performs on par or better than human experts when only audio is available, especially on challenging but important subgroups of the data: younger patients, mild cases, and female patients. Whisper's ability to recognize acoustic cues in difficult cases complements the multimodal and contextual strengths of human experts.

comment: Accepted to SMASH 2025

☆ AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs

Peize He, Zichen Wen, Yubo Wang, Yuxuan Wang, Xiaoqian Liu, Jiajie Huang, Zehui Lei, Zhuangcheng Gu, Xiangqi Jin, Jiabing Yang, Kai Li, Zhifei Liu, Weijia Li, Cunxiang Wang, Conghui He, Linfeng Zhang

Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.

comment: 26 pages, 23 figures, the code is available at \url{https://github.com/DabDans/AudioMarathon}

☆ Making Machines Sound Sarcastic: LLM-Enhanced and Retrieval-Guided Sarcastic Speech Synthesis

Zhu Li, Yuqing Zhang, Xiyuan Gao, Shekhar Nayak, Matt Coler

Sarcasm is a subtle form of non-literal language that poses significant challenges for speech synthesis due to its reliance on nuanced semantic, contextual, and prosodic cues. While existing speech synthesis research has focused primarily on broad emotional categories, sarcasm remains largely unexplored. In this paper, we propose a Large Language Model (LLM)-enhanced Retrieval-Augmented framework for sarcasm-aware speech synthesis. Our approach combines (1) semantic embeddings from a LoRA-fine-tuned LLaMA 3, which capture pragmatic incongruity and discourse-level cues of sarcasm, and (2) prosodic exemplars retrieved via a Retrieval Augmented Generation (RAG) module, which provide expressive reference patterns of sarcastic delivery. Integrated within a VITS backbone, this dual conditioning enables more natural and contextually appropriate sarcastic speech. Experiments demonstrate that our method outperforms baselines in both objective measures and subjective evaluations, yielding improvements in speech naturalness, sarcastic expressivity, and downstream sarcasm detection.

☆ Open ASR Leaderboard: Towards Reproducible and Transparent Multilingual and Long-Form Speech Recognition Evaluation ICASSP 2026

Vaibhav Srivastav, Steven Zheng, Eric Bezzam, Eustache Le Bihan, Nithin Koluguri, Piotr Żelasko, Somshubra Majumdar, Adel Moumen, Sanchit Gandhi

Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including dedicated multilingual and long-form tracks. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

comment: Submitted to ICASSP 2026; Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard; Code: https://github.com/huggingface/open_asr_leaderboard

☆ Towards Responsible Evaluation for Text-to-Speech

Yifan Yang, Hui Wang, Bing Han, Shujie Liu, Jinyu Li, Yong Qin, Xie Chen

Recent advances in text-to-speech (TTS) technology have enabled systems to produce human-indistinguishable speech, bringing benefits across accessibility, content creation, and human-computer interaction. However, current evaluation practices are increasingly inadequate for capturing the full range of capabilities, limitations, and societal implications. This position paper introduces the concept of Responsible Evaluation and argues that it is essential and urgent for the next phase of TTS development, structured through three progressive levels: (1) ensuring the faithful and accurate reflection of a model's true capabilities, with more robust, discriminative, and comprehensive objective and subjective scoring methodologies; (2) enabling comparability, standardization, and transferability through standardized benchmarks, transparent reporting, and transferable evaluation metrics; and (3) assessing and mitigating ethical risks associated with forgery, misuse, privacy violations, and security vulnerabilities. Through this concept, we critically examine current evaluation practices, identify systemic shortcomings, and propose actionable recommendations. We hope this concept of Responsible Evaluation will foster more trustworthy and reliable TTS technology and guide its development toward ethically sound and societally beneficial applications.

☆ SHANKS: Simultaneous Hearing and Thinking for Spoken Language Models

Cheng-Han Chiang, Xiaofei Wang, Linjie Li, Chung-Ching Lin, Kevin Lin, Shujie Liu, Zhendong Wang, Zhengyuan Yang, Hung-yi Lee, Lijuan Wang

Current large language models (LLMs) and spoken language models (SLMs) begin thinking and taking actions only after the user has finished their turn. This prevents the model from interacting during the user's turn and can lead to high response latency while it waits to think. Consequently, thinking after receiving the full input is not suitable for speech-to-speech interaction, where real-time, low-latency exchange is important. We address this by noting that humans naturally "think while listening." In this paper, we propose SHANKS, a general inference framework that enables SLMs to generate unspoken chain-of-thought reasoning while listening to the user input. SHANKS streams the input speech in fixed-duration chunks and, as soon as a chunk is received, generates unspoken reasoning based on all previous speech and reasoning, while the user continues speaking. SHANKS uses this unspoken reasoning to decide whether to interrupt the user and to make tool calls to complete the task. We demonstrate that SHANKS enhances real-time user-SLM interaction in two scenarios: (1) when the user is presenting a step-by-step solution to a math problem, SHANKS can listen, reason, and interrupt when the user makes a mistake, achieving 37.1% higher interruption accuracy than a baseline that interrupts without thinking; and (2) in a tool-augmented dialogue, SHANKS can complete 56.9% of the tool calls before the user finishes their turn. Overall, SHANKS moves toward models that keep thinking throughout the conversation, not only after a turn ends. Animated illustrations of Shanks can be found at https://d223302.github.io/SHANKS/

comment: Work in progress

☆ Moises-Light: Resource-efficient Band-split U-Net For Music Source Separation

Yun-Ning, Hung, Igor Pereira, Filip Korzeniowski

In recent years, significant advances have been made in music source separation, with model architectures such as dual-path modeling, band-split modules, or transformer layers achieving comparably good results. However, these models often contain a significant number of parameters, posing challenges to devices with limited computational resources in terms of training and practical application. While some lightweight models have been introduced, they generally perform worse compared to their larger counterparts. In this paper, we take inspiration from these recent advances to improve a lightweight model. We demonstrate that with careful design, a lightweight model can achieve comparable SDRs to models with up to 13 times more parameters. Our proposed model, Moises-Light, achieves competitive results in separating four musical stems on the MUSDB-HQ benchmark dataset. The proposed model also demonstrates competitive scalability when using MoisesDB as additional training data.

☆ XLSR-Kanformer: A KAN-Intergrated model for Synthetic Speech Detection

Phuong Tuan Dat, Tran Huy Dat

Recent advancements in speech synthesis technologies have led to increasingly sophisticated spoofing attacks, posing significant challenges for automatic speaker verification systems. While systems based on self-supervised learning (SSL) models, particularly the XLSR-Conformer architecture, have demonstrated remarkable performance in synthetic speech detection, there remains room for architectural improvements. In this paper, we propose a novel approach that replaces the traditional Multi-Layer Perceptron (MLP) in the XLSR-Conformer model with a Kolmogorov-Arnold Network (KAN), a powerful universal approximator based on the Kolmogorov-Arnold representation theorem. Our experimental results on ASVspoof2021 demonstrate that the integration of KAN to XLSR-Conformer model can improve the performance by 60.55% relatively in Equal Error Rate (EER) LA and DF sets, further achieving 0.70% EER on the 21LA set. Besides, the proposed replacement is also robust to various SSL architectures. These findings suggest that incorporating KAN into SSL-based models is a promising direction for advances in synthetic speech detection.

comment: Accepted to 2025 IEEE International Conference on Advanced Video and Signal-Based Surveillance

☆ Learning to Rewrite Prompts for Bootstrapping LLMs on Downstream Tasks

Qinhao Zhou, Xiang Xiang, Kun He, John E. Hopcroft

In recent years, the growing interest in Large Language Models (LLMs) has significantly advanced prompt engineering, transitioning from manual design to model-based optimization. Prompts for LLMs generally comprise two components: the \textit{instruction}, which defines the task or objective, and the \textit{input}, which is tailored to the instruction type. In natural language generation (NLG) tasks such as machine translation, the \textit{input} component is particularly critical, while the \textit{instruction} component tends to be concise. Existing prompt engineering methods primarily focus on optimizing the \textit{instruction} component for general tasks, often requiring large-parameter LLMs as auxiliary tools. However, these approaches exhibit limited applicability for tasks like machine translation, where the \textit{input} component plays a more pivotal role. To address this limitation, this paper introduces a novel prompt optimization method specifically designed for machine translation tasks. The proposed approach employs a small-parameter model trained using a back-translation-based strategy, significantly reducing training overhead for single-task optimization while delivering highly effective performance. With certain adaptations, this method can also be extended to other downstream tasks.

☆ Utilizing Information Theoretic Approach to Study Cochlear Neural Degeneration

Ahsan J. Cheema, Sunil Puria

Hidden hearing loss, or cochlear neural degeneration (CND), disrupts suprathreshold auditory coding without affecting clinical thresholds, making it difficult to diagnose. We present an information-theoretic framework to evaluate speech stimuli that maximally reveal CND by quantifying mutual information (MI) loss between inner hair cell (IHC) receptor potentials and auditory nerve fiber (ANF) responses and acoustic input and ANF responses. Using a phenomenological auditory model, we simulated responses to 50 CVC words under clean, time-compressed, reverberant, and combined conditions across different presentation levels, with systematically varied survival of low-, medium-, and high-spontaneous-rate fibers. MI was computed channel-wise between IHC and ANF responses and integrated across characteristic frequencies. Information loss was defined relative to a normal-hearing baseline. Results demonstrate progressive MI loss with increasing CND, most pronounced for time-compressed speech, while reverberation produced comparatively smaller effects. These findings identify rapid, temporally dense speech as optimal probes for CND, informing the design of objective clinical diagnostics while revealing problems associated with reverberation as a probe.

☆ Pitch Estimation With Mean Averaging Smoothed Product Spectrum And Musical Consonance Evaluation Using MASP

Murat Yasar Baskin

This study introduces Mean Averaging Smoothed Product (MASP) Spectrum, which is a modified version of the Harmonic Product Spectrum, designed to enhance pitch estimation for many algorithm-wise deceptive frequency spectra that still lead clear pitches, for both harmonic and inharmonic cases. By introducing a global mean based smoothing for spectrum, the MASP algorithm diminishes the unwanted sensitivity of HPS for spectra with missing partials. The method exhibited robust pitch estimations consistent with perceptual expectations. Motivated upon the strong correlation between consonance and periodicity, the same algorithm is extended and, with the proposition of a harmonicity measure (H), used to evaluate musical consonance for two and three tones; yielding consonance hierarchies that align with perception and practice of music theory. These findings suggest that perception of pitch and consonance may share a similar underlying mechanism that depend on spectrum.

☆ Benchmarking Fake Voice Detection in the Fake Voice Generation Arms Race

Xutao Mao, Ke Li, Cameron Baird, Ezra Xuanru Tao, Dan Lin

As advances in synthetic voice generation accelerate, an increasing variety of fake voice generators have emerged, producing audio that is often indistinguishable from real human speech. This evolution poses new and serious threats across sectors where audio recordings serve as critical evidence. Although fake voice detectors are also advancing, the arms race between fake voice generation and detection has become more intense and complex. In this work, we present the first large-scale, cross-domain evaluation of fake voice detectors, benchmarking 8 state-of-the-art models against datasets synthesized by 20 different fake voice generation systems. To the best of our knowledge, this is the most comprehensive cross-domain assessment conducted to date. Our study reveals substantial security vulnerabilities in current fake voice detection systems, underscoring critical gaps in their real-world robustness. To advance the field, we propose a unified and effective metric that consolidates the diverse and often inconsistent evaluation criteria previously used across different studies. This metric enables standardized, straightforward comparisons of the robustness of fake voice detectors. We conclude by offering actionable recommendations for building more resilient fake voice detection technologies, with the broader goal of reinforcing the foundations of AI security and trustworthiness.

☆ BACHI: Boundary-Aware Symbolic Chord Recognition Through Masked Iterative Decoding on Pop and Classical Music

Mingyang Yao, Ke Chen, Shlomo Dubnov, Taylor Berg-Kirkpatrick

Automatic chord recognition (ACR) via deep learning models has gradually achieved promising recognition accuracy, yet two key challenges remain. First, prior work has primarily focused on audio-domain ACR, while symbolic music (e.g., score) ACR has received limited attention due to data scarcity. Second, existing methods still overlook strategies that are aligned with human music analytical practices. To address these challenges, we make two contributions: (1) we introduce POP909-CL, an enhanced version of POP909 dataset with tempo-aligned content and human-corrected labels of chords, beats, keys, and time signatures; and (2) We propose BACHI, a symbolic chord recognition model that decomposes the task into different decision steps, namely boundary detection and iterative ranking of chord root, quality, and bass (inversion). This mechanism mirrors the human ear-training practices. Experiments demonstrate that BACHI achieves state-of-the-art chord recognition performance on both classical and pop music benchmarks, with ablation studies validating the effectiveness of each module.

comment: Under review

☆ SALAD-VAE: Semantic Audio Compression with Language-Audio Distillation ICASSP 2026

Sebastian Braun, Hannes Gamper, Dimitra Emmanouilidou

Modern generative and multimodal models increasingly rely on compact latent representations that trade and balance semantic richness with high-fidelity reconstruction. We introduce SALAD-VAE, a continuous and highly compact semantic Audio Variational Autoencoder, which operates in the frequency domain and achieves state-of-the-art compression with very low latent frame rate (7.8 Hz) while surfacing semantic structure and producing high audio quality. We enhance the standard VAE semantic losses and augmentation, specifically contrastive learning and CLAP-based embedding distillation, enabling it to generalize across diverse audio domains. With a significantly less computational complex architecture than comparable state-of-the-art VAEs, SALAD-VAE matches their reconstruction quality while it consistently outperforms them on a wide range of classification benchmarks. Furthermore, the proposed additional loss function provides a trained CLAP projection layer, which can be used zero-shot audio captioning and classification matching pretrained CLAP audio-text embeddings.

comment: submitted to ICASSP 2026

☆ Can Speech LLMs Think while Listening?

Yi-Jen Shih, Desh Raj, Chunyang Wu, Wei Zhou, SK Bong, Yashesh Gaur, Jay Mahadeokar, Ozlem Kalinli, Mike Seltzer

Recent advances in speech large language models (speech LLMs) have enabled seamless spoken interactions, but these systems still struggle with complex reasoning tasks. Previously, chain-of-thought (CoT) prompting or fine-tuning has been to shown to significantly improve the reasoning abilities of text-based LLMs. In this work, we investigate the effect of CoT fine-tuning for multi-stream speech LLMs, demonstrating that reasoning in text space improves the accuracy of speech LLMs by 2.4x, on average, over a suite of spoken reasoning tasks. Beyond accuracy, the latency of the spoken response is a crucial factor for interacting with voice-based agents. Inspired by the human behavior of "thinking while listening," we propose methods to reduce the additional latency from reasoning by allowing the model to start reasoning before the user query has ended. To achieve this, we introduce an entropy-based metric, "question completeness," which acts as an indicator to guide the model on the optimal time to start reasoning. This method provides greater control over the accuracy-latency trade-off compared with heuristic-based approaches and, under equivalent latency conditions, yields a 4% accuracy gain on ARC-Easy. Finally, we use Direct Preference Optimization (DPO) on preference data created using rejection sampling to push the accuracy-latency pareto frontier further, resulting in a 70% reduction in latency without loss in accuracy.

☆ LASER: An LLM-based ASR Scoring and Evaluation Rubric EMNLP 2025

Amruta Parulekar, Preethi Jyothi

Standard ASR evaluation metrics like Word Error Rate (WER) tend to unfairly penalize morphological and syntactic nuances that do not significantly alter sentence semantics. We introduce an LLM-based scoring rubric LASER that leverages state-of-the-art LLMs' in-context learning abilities to learn from prompts with detailed examples. Hindi LASER scores using Gemini 2.5 Pro achieved a very high correlation score of 94% with human annotations. Hindi examples in the prompt were also effective in analyzing errors in other Indian languages such as Marathi, Kannada and Malayalam. We also demonstrate how a smaller LLM like Llama 3 can be finetuned on word-pair examples derived from reference and ASR predictions to predict what kind of penalty should be applied with close to 89% accuracy.

comment: Accepted to EMNLP 2025

☆ Look before Transcription: End-to-End SlideASR with Visually-Anchored Policy Optimization

Rui Hu, Delai Qiu, Yining Wang, Shengping Liu, Jitao Sang

Automatic speech recognition (ASR) systems often struggle with domain-specific terminology, especially in specialized settings such as academic lectures. To address this, we define the SlideASR task, which leverages the rich visual information from presentation slides to improve transcription accuracy. Existing pipeline methods for this task tend to be complex and underperform. Although omni-modal large language models (OLLMs) provide a promising end-to-end framework, they frequently fail in practice by degenerating into simple optical character recognition (OCR) systems. To overcome this, we propose Visually-Anchored Policy Optimization (VAPO), a novel post-training method designed to control the model's reasoning process. Drawing on the Chain-of-Thought reasoning paradigm, VAPO enforces a structured "Look before Transcription" procedure using a format. Specifically, the model first performs OCR on the slide content within the think step, then generates the transcription by referencing this recognized visual information in the answer step. This reasoning process is optimized via reinforcement learning with four distinct rewards targeting format compliance, OCR accuracy, ASR quality, and visual anchoring consistency. To support further research, we construct SlideASR-Bench, a new entity-rich benchmark consisting of a synthetic dataset for training and testing, and a challenging real-world set for evaluation. Extensive experiments demonstrate that VAPO significantly improves recognition of domain-specific terms, establishing an effective end-to-end paradigm for SlideASR.

♻ ☆ LaunchpadGPT: Language Model as Music Visualization Designer on Launchpad

Siting Xu, Yolo Yunlong Tang, Feng Zheng

Launchpad is a musical instrument that allows users to create and perform music by pressing illuminated buttons. To assist and inspire the design of the Launchpad light effect, and provide a more accessible approach for beginners to create music visualization with this instrument, we proposed the LaunchpadGPT model to generate music visualization designs on Launchpad automatically. Based on the language model with excellent generation ability, our proposed LaunchpadGPT takes an audio piece of music as input and outputs the lighting effects of Launchpad-playing in the form of a video (Launchpad-playing video). We collect Launchpad-playing videos and process them to obtain music and corresponding video frame of Launchpad-playing as prompt-completion pairs, to train the language model. The experiment result shows the proposed method can create better music visualization than random generation methods and hold the potential for a broader range of music visualization applications. Our code is available at https://github.com/yunlong10/LaunchpadGPT/.

comment: Accepted to International Computer Music Conference (ICMC) 2023

♻ ☆ AbsoluteNet: A Deep Learning Neural Network to Classify Cerebral Hemodynamic Responses of Auditory Processing

Behtom Adeli, John Mclinden, Pankaj Pandey, Ming Shao, Yalda Shahriari

In recent years, deep learning (DL) approaches have demonstrated promising results in decoding hemodynamic responses captured by functional near-infrared spectroscopy (fNIRS), particularly in the context of brain-computer interface (BCI) applications. This work introduces AbsoluteNet, a novel deep learning architecture designed to classify auditory event-related responses recorded using fNIRS. The proposed network is built upon principles of spatio-temporal convolution and customized activation functions. Our model was compared against several models, namely fNIRSNET, MDNN, DeepConvNet, and ShallowConvNet. The results showed that AbsoluteNet outperforms existing models, reaching 87.0% accuracy, 84.8% sensitivity, and 89.2% specificity in binary classification, surpassing fNIRSNET, the second-best model, by 3.8% in accuracy. These findings underscore the effectiveness of our proposed deep learning model in decoding hemodynamic responses related to auditory processing and highlight the importance of spatio-temporal feature aggregation and customized activation functions to better fit fNIRS dynamics.

♻ ☆ The Sound of Syntax: Finetuning and Comprehensive Evaluation of Language Models for Speech Pathology EMNLP 2025

Fagun Patel, Duc Q. Nguyen, Sang T. Truong, Jody Vaynshtok, Sanmi Koyejo, Nick Haber

According to the U.S. National Institutes of Health, more than 3.4 million children experience speech disorders that require clinical intervention. The number of speech-language pathologists (SLPs) is roughly 20 times fewer than the number of affected children, highlighting a significant gap in children's care and a pressing need for technological support that improves the productivity of SLPs. State-of-the-art multimodal language models (MLMs) show promise for supporting SLPs, but their use remains underexplored largely due to a limited understanding of their performance in high-stakes clinical settings. To address this gap, we collaborate with domain experts to develop a taxonomy of real-world use cases of MLMs in speech-language pathologies. Building on this taxonomy, we introduce the first comprehensive benchmark for evaluating MLM across five core use cases, each containing 1,000 manually annotated data points. This benchmark includes robustness and sensitivity tests under various settings, including background noise, speaker gender, and accent. Our evaluation of 15 state-of-the-art MLMs reveals that no single model consistently outperforms others across all tasks. Notably, we find systematic disparities, with models performing better on male speakers, and observe that chain-of-thought prompting can degrade performance on classification tasks with large label spaces and narrow decision boundaries. Furthermore, we study fine-tuning MLMs on domain-specific data, achieving improvements of over 10\% compared to base models. These findings highlight both the potential and limitations of current MLMs for speech-language pathology applications, underscoring the need for further research and targeted development.

comment: EMNLP 2025 Oral Presentation

♻ ☆ A Differentiable Alignment Framework for Sequence-to-Sequence Modeling via Optimal Transport

Yacouba Kaloga, Shashi Kumar, Petr Motlicek, Ina Kodrasi

Accurate sequence-to-sequence (seq2seq) alignment is critical for applications like medical speech analysis and language learning tools relying on automatic speech recognition (ASR). State-of-the-art end-to-end (E2E) ASR systems, such as the Connectionist Temporal Classification (CTC) and transducer-based models, suffer from peaky behavior and alignment inaccuracies. In this paper, we propose a novel differentiable alignment framework based on one-dimensional optimal transport, enabling the model to learn a single alignment and perform ASR in an E2E manner. We introduce a pseudo-metric, called Sequence Optimal Transport Distance (SOTD), over the sequence space and discuss its theoretical properties. Based on the SOTD, we propose Optimal Temporal Transport Classification (OTTC) loss for ASR and contrast its behavior with CTC. Experimental results on the TIMIT, AMI, and LibriSpeech datasets show that our method considerably improves alignment performance compared to CTC and the more recently proposed Consistency-Regularized CTC, though with a trade-off in ASR performance. We believe this work opens new avenues for seq2seq alignment research, providing a solid foundation for further exploration and development within the community.

♻ ☆ Enhancing Few-shot Keyword Spotting Performance through Pre-Trained Self-supervised Speech Models

Alican Gok, Oguzhan Buyuksolak, Osman Erman Okman, Murat Saraclar

Keyword Spotting plays a critical role in enabling hands-free interaction for battery-powered edge devices. Few-Shot Keyword Spotting (FS-KWS) addresses the scalability and adaptability challenges of traditional systems by enabling recognition of custom keywords with only a few examples. However, existing FS-KWS systems achieve subpar accuracy at desirable false acceptance rates, particularly in resource-constrained edge environments. To address these issues, we propose a training scheme that leverages self-supervised learning models for robust feature extraction, dimensionality reduction, and knowledge distillation. The teacher model, based on Wav2Vec 2.0 is trained using Sub-center ArcFace loss, which enhances inter-class separability and intra-class compactness. To enable efficient deployment on edge devices, we introduce attention-based dimensionality reduction and train a standard lightweight ResNet15 student model. We evaluate the proposed approach on the English portion of the Multilingual Spoken Words Corpus (MSWC) and the Google Speech Commands (GSC) datasets. Notably, the proposed training method improves the 10-shot classification accuracy from 33.4% to 74.1% on 11 classes at 1% false alarm accuracy on the GSC dataset, thus making it significantly better-suited for a real use case scenario.

comment: Submitted to IEEE Signal Processing Letters, 5 pages, 3 figures

♻ ☆ Token-based Audio Inpainting via Discrete Diffusion

Tali Dror, Iftach Shoham, Moshe Buchris, Oren Gal, Haim Permuter, Gilad Katz, Eliya Nachmani

Audio inpainting seeks to restore missing segments in degraded recordings. Previous diffusion-based methods exhibit impaired performance when the missing region is large. We introduce the first approach that applies discrete diffusion over tokenized music representations from a pre-trained audio tokenizer, enabling stable and semantically coherent restoration of long gaps. Our method further incorporates two training approaches: a derivative-based regularization loss that enforces smooth temporal dynamics, and a span-based absorbing transition that provides structured corruption during diffusion. Experiments on the MusicNet and MAESTRO datasets with gaps up to 750 ms show that our approach consistently outperforms strong baselines across range of gap lengths, for gaps of 150 ms and above. This work advances musical audio restoration and introduces new directions for discrete diffusion model training. Audio examples of our proposed method can be found at https://iftach21.github.io/.

♻ ☆ Enhancing Situational Awareness in Wearable Audio Devices Using a Lightweight Sound Event Localization and Detection System

Jun-Wei Yeow, Ee-Leng Tan, Santi Peksi, Zhen-Ting Ong, Woon-Seng Gan

Wearable audio devices with active noise control (ANC) enhance listening comfort but often at the expense of situational awareness. However, this auditory isolation may mask crucial environmental cues, posing significant safety risks. To address this, we propose an environmental intelligence framework that combines Acoustic Scene Classification (ASC) with Sound Event Localization and Detection (SELD). Our system first employs a lightweight ASC model to infer the current environment. The scene prediction then dynamically conditions a SELD network, tuning its sensitivity to detect and localize sounds that are most salient to the current context. On simulated headphone data, the proposed ASC-conditioned SELD system demonstrates improved spatial intelligence over a conventional baseline. This work represents a crucial step towards creating intelligent hearables that can deliver crucial environmental information, fostering a safer and more context-aware listening experience.

comment: Accepted and presented at the 2025 AES International Conference on Headphone Technology

♻ ☆ Good practices for evaluation of synthesized speech

Erica Cooper, Sébastien Le Maguer, Esther Klabbers, Junichi Yamagishi

This document is provided as a guideline for reviewers of papers about speech synthesis. We outline some best practices and common pitfalls for papers about speech synthesis, with a particular focus on evaluation. We also recommend that reviewers check the guidelines for authors written in the paper kit and consider those as reviewing criteria as well. This is intended to be a living document, and it will be updated as we receive comments and feedback from readers. We note that this document is meant to provide guidance only, and that reviewers should ultimately use their own discretion when evaluating papers.

♻ ☆ Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation

Haorui He, Zengqiang Shang, Chaoren Wang, Xuyuan Li, Yicheng Gu, Hua Hua, Liwei Liu, Chen Yang, Jiaqi Li, Peiyang Shi, Yuancheng Wang, Kai Chen, Pengyuan Zhang, Zhizheng Wu

Recent advancements in speech generation have been driven by large-scale training datasets. However, current models struggle to capture the spontaneity and variability inherent in real-world human speech, as they are primarily trained on audio-book datasets limited to formal, read-aloud speaking styles. To address this limitation, we introduce Emilia-Pipe, an open-source preprocessing pipeline designed to extract high-quality training data from valuable yet under-explored in-the-wild sources that capture spontaneous human speech in real-world contexts. Using Emilia-Pipe, we construct Emilia, which comprises over 101k hours of speech across six languages: English, Chinese, German, French, Japanese, and Korean. Furthermore, we expand Emilia to Emilia-Large, a dataset exceeding 216k hours, making it one of the largest open-source speech generation resources available. Extensive experiments show that Emilia-trained models produce markedly more spontaneous, human-like speech than those trained on traditional audio-book datasets, while matching their intelligibility. These models better capture diverse speaker timbres and the full spectrum of real-world conversational styles. Our work also highlights the importance of scaling dataset size for advancing speech generation performance and validates the effectiveness of Emilia for both multilingual and crosslingual speech generation tasks.

comment: Full version of arXiv:2407.05361, dataset is available at: https://huggingface.co/datasets/amphion/Emilia-Dataset

♻ ☆ PredGen: Accelerated Inference of Large Language Models through Input-Time Speculation for Real-Time Speech Interaction

Shufan Li, Aditya Grover

Large Language Models (LLMs) are widely used in real-time voice chat applications, typically in combination with text-to-speech (TTS) systems to generate audio responses. However, their large size often leads to noticeable latency between the end of user input and the start of audio output, resulting in suboptimal user experiences. This latency is particularly evident when LLMs are deployed as single-user voice assistants on consumer-grade hardware with limited computing capacity. We discovered that this latency is primarily dominated by the time it takes for the LLMs to generate the first sentence, which is required as input by the TTS systems that synthesize audio responses on a sentence-by-sentence basis. To address this bottleneck, we propose Predictive Generation (PredGen), a novel framework that mitigates-or even eliminates-this delay through speculative decoding at input time. PredGen generates candidate responses while the user is still speaking, enabling the system to begin TTS processing with minimal delay. Simulated experiments on the Lmsys and MT-Bench datasets show that the proposed method can effectively reduce the latency by around 2x across a wide range of use cases, while incurring only minimal additional computation cost at input time-computation that would otherwise go unused.

comment: 16 pages,4 figures

♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS

Seungheon Doh, Keunwoo Choi, Juhan Nam

While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.

comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)

♻ ☆ Self-Supervised Speech Quality Assessment (S3QA): Leveraging Speech Foundation Models for a Scalable Speech Quality Metric

Mattson Ogg, Caitlyn Bishop, Han Yi, Sarah Robinson

Methods for automatically assessing speech quality in real world environments are critical for developing robust human language technologies and assistive devices. Behavioral ratings provided by human raters (e.g., mean opinion scores; MOS) are considered the gold standard, but they are susceptible to variability between individual raters, cannot easily be generalized across corpora, and are labor-intensive to collect, thus limiting the acoustic challenges they can quantify. Here, we present a new, scalable method for automatically assessing speech quality: the self-supervised speech quality assessment (S3QA) model. First, we manipulated high quality utterances from multiple speech corpora, using a wide range of acoustic challenges intended to emulate common sources of quality degradation in the real-world: frequency filtering, reverberation, background noise, and digital compression. Second, we leveraged an existing, pre-trained speech foundation model, WavLM, to computationally derive a self-supervised training target that quantified speech degradation using the cosine distance between the clean and degraded versions of each utterance in the embedding space. Next, we trained a transformer-based model to predict these cosine distances, given only the degraded versions of the utterances. Finally, the trained model was evaluated on unseen test corpora of synthetic mixtures, NISQA, and VOiCES. We show that the S3QA model trained on this task accurately predicts degradation cosine distances across a wide range challenging acoustic conditions and is aligned with both behavioral ratings (MOS), speech technology performance (automatic speech recognition) and other important features of the held-out data (e.g., microphone distances). This model provides an automated, scalable method for assessing speech quality across a wide range of acoustic challenges.

comment: 1 table, seven figures, thirteen pages

♻ ☆ TalkPlayData 2: An Agentic Synthetic Data Pipeline for Multimodal Conversational Music Recommendation

Keunwoo Choi, Seungheon Doh, Juhan Nam

We present TalkPlayData 2, a synthetic dataset for multimodal conversational music recommendation generated by an agentic data pipeline. In the proposed pipeline, multiple large language model (LLM) agents are created under various roles with specialized prompts and access to different parts of information, and the chat data is acquired by logging the conversation between the Listener LLM and the Recsys LLM. To cover various conversation scenarios, for each conversation, the Listener LLM is conditioned on a finetuned conversation goal. Finally, all the LLMs are multimodal with audio and images, allowing a simulation of multimodal recommendation and conversation. In the LLM-as-a-judge and subjective evaluation experiments, TalkPlayData 2 achieved the proposed goal in various aspects related to training a generative recommendation model for music. TalkPlayData 2 and its generation code are released at https://talkpl.ai/talkplaydata2.

comment: 2025-10-08: updating the stat table with the latest numbers. updated the abstract per the latest license terms

♻ ☆ Evaluating Sound Similarity Metrics for Differentiable, Iterative Sound-Matching

Amir Salimi, Abram Hindle, Osmar R. Zaiane

Manual sound design with a synthesizer is inherently iterative: an artist compares the synthesized output to a mental target, adjusts parameters, and repeats until satisfied. Iterative sound-matching automates this workflow by continually programming a synthesizer under the guidance of a loss function (or similarity measure) toward a target sound. Prior comparisons of loss functions have typically favored one metric over another, but only within narrow settings: limited synthesis methods, few loss types, often without blind listening tests. This leaves open the question of whether a universally optimal loss exists, or the choice of loss remains a creative decision conditioned on the synthesis method and the sound designer's preference. We propose differentiable iterative sound-matching as the natural extension of the available literature, since it combines the manual approach to sound design with modern advances in machine learning. To analyze the variability of loss function performance across synthesizers, we implemented a mix of four novel and established differentiable loss functions, and paired them with differentiable subtractive, additive, and AM synthesizers. For each of the sixteen synthesizer--loss combinations, we ran 300 randomized sound-matching trials. Performance was measured using parameter differences, spectrogram-distance metrics, and manually assigned listening scores. We observed a moderate level of consistency among the three performance measures. Our post-hoc analysis shows that the loss function performance is highly dependent on the synthesizer. These findings underscore the value of expanding the scope of sound-matching experiments and developing new similarity metrics tailored to specific synthesis techniques rather than pursuing one-size-fits-all solutions.

☆ Modulation Discovery with Differentiable Digital Signal Processing SP

Christopher Mitcheltree, Hao Hao Tan, Joshua D. Reiss

Modulations are a critical part of sound design and music production, enabling the creation of complex and evolving audio. Modern synthesizers provide envelopes, low frequency oscillators (LFOs), and more parameter automation tools that allow users to modulate the output with ease. However, determining the modulation signals used to create a sound is difficult, and existing sound-matching / parameter estimation systems are often uninterpretable black boxes or predict high-dimensional framewise parameter values without considering the shape, structure, and routing of the underlying modulation curves. We propose a neural sound-matching approach that leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover the modulations present in a sound. We demonstrate the effectiveness of our approach on highly modulated synthetic and real audio samples, its applicability to different DDSP synth architectures, and investigate the trade-off it incurs between interpretability and sound-matching accuracy. We make our code and audio samples available and provide the trained DDSP synths in a VST plugin.

comment: Accepted to WASPAA 2025 (best paper award candidate). Code, audio samples, and plugins can be found at https://christhetree.github.io/mod_discovery/

☆ TokenChain: A Discrete Speech Chain via Semantic Token Modeling ICASSP

Mingxuan Wang, Satoshi Nakamura

Machine Speech Chain, simulating the human perception-production loop, proves effective in jointly improving ASR and TTS. We propose TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS: an autoregressive text-to-semantic model co-trained with ASR and a masked-generative semantic-to-acoustic model for synthesis only. End-to-end feedback across the text interface is enabled with straight-through argmax/Gumbel-Softmax and balanced with supervised ASR via dynamic weight averaging. Ablations examine optimal temperature schedules for in- and cross-domain transfer. Evaluation reveals TokenChain surpasses baseline accuracy 2-6 epochs earlier and yields 5-13% lower equal-epoch error with stable T2S on LibriSpeech, and reduces relative ASR WER by 56% and T2S WER by 31% on TED-LIUM with minimal forgetting, showing that chain learning remains effective with token interfaces and models.

comment: 5 pages, 3 figures. Submitted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026

☆ EmoHRNet: High-Resolution Neural Network Based Speech Emotion Recognition

Akshay Muppidi, Martin Radfar

Speech emotion recognition (SER) is pivotal for enhancing human-machine interactions. This paper introduces "EmoHRNet", a novel adaptation of High-Resolution Networks (HRNet) tailored for SER. The HRNet structure is designed to maintain high-resolution representations from the initial to the final layers. By transforming audio samples into spectrograms, EmoHRNet leverages the HRNet architecture to extract high-level features. EmoHRNet's unique architecture maintains high-resolution representations throughout, capturing both granular and overarching emotional cues from speech signals. The model outperforms leading models, achieving accuracies of 92.45% on RAVDESS, 80.06% on IEMOCAP, and 92.77% on EMOVO. Thus, we show that EmoHRNet sets a new benchmark in the SER domain.

☆ ECTSpeech: Enhancing Efficient Speech Synthesis via Easy Consistency Tuning

Tao Zhu, Yinfeng Yu, Liejun Wang, Fuchun Sun, Wendong Zheng

Diffusion models have demonstrated remarkable performance in speech synthesis, but typically require multi-step sampling, resulting in low inference efficiency. Recent studies address this issue by distilling diffusion models into consistency models, enabling efficient one-step generation. However, these approaches introduce additional training costs and rely heavily on the performance of pre-trained teacher models. In this paper, we propose ECTSpeech, a simple and effective one-step speech synthesis framework that, for the first time, incorporates the Easy Consistency Tuning (ECT) strategy into speech synthesis. By progressively tightening consistency constraints on a pre-trained diffusion model, ECTSpeech achieves high-quality one-step generation while significantly reducing training complexity. In addition, we design a multi-scale gate module (MSGate) to enhance the denoiser's ability to fuse features at different scales. Experimental results on the LJSpeech dataset demonstrate that ECTSpeech achieves audio quality comparable to state-of-the-art methods under single-step sampling, while substantially reducing the model's training cost and complexity.

comment: Accepted for publication by Proceedings of the 2025 ACM Multimedia Asia Conference(MMAsia '25)

☆ Segment-Factorized Full-Song Generation on Symbolic Piano Music NeurIPS 2025

Ping-Yi Chen, Chih-Pin Tan, Yi-Hsuan Yang

We propose the Segmented Full-Song Model (SFS) for symbolic full-song generation. The model accepts a user-provided song structure and an optional short seed segment that anchors the main idea around which the song is developed. By factorizing a song into segments and generating each one through selective attention to related segments, the model achieves higher quality and efficiency compared to prior work. To demonstrate its suitability for human-AI interaction, we further wrap SFS into a web application that enables users to iteratively co-create music on a piano roll with customizable structures and flexible ordering.

comment: Accepted to the 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: AI for Music

☆ LARA-Gen: Enabling Continuous Emotion Control for Music Generation Models via Latent Affective Representation Alignment

Jiahao Mei, Xuenan Xu, Zeyu Xie, Zihao Zheng, Ye Tao, Yue Ding, Mengyue Wu

Recent advances in text-to-music models have enabled coherent music generation from text prompts, yet fine-grained emotional control remains unresolved. We introduce LARA-Gen, a framework for continuous emotion control that aligns the internal hidden states with an external music understanding model through Latent Affective Representation Alignment (LARA), enabling effective training. In addition, we design an emotion control module based on a continuous valence-arousal space, disentangling emotional attributes from textual content and bypassing the bottlenecks of text-based prompting. Furthermore, we establish a benchmark with a curated test set and a robust Emotion Predictor, facilitating objective evaluation of emotional controllability in music generation. Extensive experiments demonstrate that LARA-Gen achieves continuous, fine-grained control of emotion and significantly outperforms baselines in both emotion adherence and music quality. Generated samples are available at https://nieeim.github.io/LARA-Gen/.

☆ FoleyGRAM: Video-to-Audio Generation with GRAM-Aligned Multimodal Encoders IJCNN 2025

Riccardo Fosco Gramaccioni, Christian Marinoni, Eleonora Grassucci, Giordano Cicchetti, Aurelio Uncini, Danilo Comminiello

In this work, we present FoleyGRAM, a novel approach to video-to-audio generation that emphasizes semantic conditioning through the use of aligned multimodal encoders. Building on prior advancements in video-to-audio generation, FoleyGRAM leverages the Gramian Representation Alignment Measure (GRAM) to align embeddings across video, text, and audio modalities, enabling precise semantic control over the audio generation process. The core of FoleyGRAM is a diffusion-based audio synthesis model conditioned on GRAM-aligned embeddings and waveform envelopes, ensuring both semantic richness and temporal alignment with the corresponding input video. We evaluate FoleyGRAM on the Greatest Hits dataset, a standard benchmark for video-to-audio models. Our experiments demonstrate that aligning multimodal encoders using GRAM enhances the system's ability to semantically align generated audio with video content, advancing the state of the art in video-to-audio synthesis.

comment: Acepted at IJCNN 2025

☆ StereoSync: Spatially-Aware Stereo Audio Generation from Video IJCNN 2025

Christian Marinoni, Riccardo Fosco Gramaccioni, Kazuki Shimada, Takashi Shibuya, Yuki Mitsufuji, Danilo Comminiello

Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.

comment: Accepted at IJCNN 2025

☆ Data-efficient Targeted Token-level Preference Optimization for LLM-based Text-to-Speech

Rikuto Kotoge, Yuichi Sasaki

Aligning text-to-speech (TTS) system outputs with human feedback through preference optimization has been shown to effectively improve the robustness and naturalness of language model-based TTS models. Current approaches primarily require paired desirable and undesirable samples at the utterance level. However, such pairs are often limited in TTS output data, and utterance-level formulation prevents fine-grained token-level optimization needed for accurate pronunciation alignment. In this study, we propose TKTO that eliminates the need for paired data, enabling a more data-efficient training paradigm, and directly targets token-level units, automatically providing fine-grained alignment signals without token-level annotations. TKTO improves the challenging Japanese TTS accuracy by 39% and reduces CER by 54%, automatically assigning 12.8 times stronger reward to targeted tokens.

☆ EMORL-TTS: Reinforcement Learning for Fine-Grained Emotion Control in LLM-based TTS ICASSP 2026

Haoxun Li, Yu Liu, Yuqing Sun, Hanlei Shi, Leyuan Qu, Taihao Li

Recent LLM-based TTS systems achieve strong quality and zero-shot ability, but lack fine-grained emotional control due to their reliance on discrete speech tokens. Existing approaches either limit emotions to categorical labels or cannot generalize to LLM-based architectures. We propose EMORL-TTS (Fine-grained Emotion-controllable TTS with Reinforcement Learning), a framework that unifies global intensity control in the VAD space with local emphasis regulation. Our method combines supervised fine-tuning with reinforcement learning guided by task-specific rewards for emotion category, intensity, and emphasis. Moreover, we further investigate how emphasis placement modulates fine-grained emotion intensity. Experiments show that EMORL-TTS improves emotion accuracy, intensity differentiation, and emphasis clarity, while preserving synthesis quality comparable to strong LLM-based baselines.

comment: Under review for ICASSP 2026

☆ Transcribing Rhythmic Patterns of the Guitar Track in Polyphonic Music SP

Aleksandr Lukoianov, Anssi Klapuri

Whereas chord transcription has received considerable attention during the past couple of decades, far less work has been devoted to transcribing and encoding the rhythmic patterns that occur in a song. The topic is especially relevant for instruments such as the rhythm guitar, which is typically played by strumming rhythmic patterns that repeat and vary over time. However, in many cases one cannot objectively define a single "right" rhythmic pattern for a given song section. To create a dataset with well-defined ground-truth labels, we asked expert musicians to transcribe the rhythmic patterns in 410 popular songs and record cover versions where the guitar tracks followed those transcriptions. To transcribe the strums and their corresponding rhythmic patterns, we propose a three-step framework. Firstly, we perform approximate stem separation to extract the guitar part from the polyphonic mixture. Secondly, we detect individual strums within the separated guitar audio, using a pre-trained foundation model (MERT) as a backbone. Finally, we carry out a pattern-decoding process in which the transcribed sequence of guitar strums is represented by patterns drawn from an expert-curated vocabulary. We show that it is possible to transcribe the rhythmic patterns of the guitar track in polyphonic music with quite high accuracy, producing a representation that is human-readable and includes automatically detected bar lines and time signature markers. We perform ablation studies and error analysis and propose a set of evaluation metrics to assess the accuracy and readability of the predicted rhythmic pattern sequence.

comment: Accepted to WASPAA 2025

☆ MSF-SER: Enriching Acoustic Modeling with Multi-Granularity Semantics for Speech Emotion Recognition ICASSP 2026

Haoxun Li, Yuqing Sun, Hanlei Shi, Yu Liu, Leyuan Qu, Taihao Li

Continuous dimensional speech emotion recognition captures affective variation along valence, arousal, and dominance, providing finer-grained representations than categorical approaches. Yet most multimodal methods rely solely on global transcripts, leading to two limitations: (1) all words are treated equally, overlooking that emphasis on different parts of a sentence can shift emotional meaning; (2) only surface lexical content is represented, lacking higher-level interpretive cues. To overcome these issues, we propose MSF-SER (Multi-granularity Semantic Fusion for Speech Emotion Recognition), which augments acoustic features with three complementary levels of textual semantics--Local Emphasized Semantics (LES), Global Semantics (GS), and Extended Semantics (ES). These are integrated via an intra-modal gated fusion and a cross-modal FiLM-modulated lightweight Mixture-of-Experts (FM-MOE). Experiments on MSP-Podcast and IEMOCAP show that MSF-SER consistently improves dimensional prediction, demonstrating the effectiveness of enriched semantic fusion for SER.

comment: Under review for ICASSP 2026

☆ Sparse deepfake detection promotes better disentanglement

Antoine Teissier, Marie Tahon, Nicolas Dugué, Aghilas Sini

Due to the rapid progress of speech synthesis, deepfake detection has become a major concern in the speech processing community. Because it is a critical task, systems must not only be efficient and robust, but also provide interpretable explanations. Among the different approaches for explainability, we focus on the interpretation of latent representations. In such paper, we focus on the last layer of embeddings of AASIST, a deepfake detection architecture. We use a TopK activation inspired by SAEs on this layer to obtain sparse representations which are used in the decision process. We demonstrate that sparse deepfake detection can improve detection performance, with an EER of 23.36% on ASVSpoof5 test set, with 95% of sparsity. We then show that these representations provide better disentanglement, using completeness and modularity metrics based on mutual information. Notably, some attacks are directly encoded in the latent space.

☆ Sci-Phi: A Large Language Model Spatial Audio Descriptor

Xilin Jiang, Hannes Gamper, Sebastian Braun

Acoustic scene perception involves describing the type of sounds, their timing, their direction and distance, as well as their loudness and reverberation. While audio language models excel in sound recognition, single-channel input fundamentally limits spatial understanding. This work presents Sci-Phi, a spatial audio large language model with dual spatial and spectral encoders that estimates a complete parameter set for all sound sources and the surrounding environment. Learning from over 4,000 hours of synthetic first-order Ambisonics recordings including metadata, Sci-Phi enumerates and describes up to four directional sound sources in one pass, alongside non-directional background sounds and room characteristics. We evaluate the model with a permutation-invariant protocol and 15 metrics covering content, location, timing, loudness, and reverberation, and analyze its robustness across source counts, signal-to-noise ratios, reverberation levels, and challenging mixtures of acoustically, spatially, or temporally similar sources. Notably, Sci-Phi generalizes to real room impulse responses with only minor performance degradation. Overall, this work establishes the first audio LLM capable of full spatial-scene description, with strong potential for real-world deployment. Demo: https://sci-phi-audio.github.io/demo

♻ ☆ CL-UZH submission to the NIST SRE 2024 Speaker Recognition Evaluation

Aref Farhadipour, Shiran Liu, Masoumeh Chapariniya, Valeriia Vyshnevetska, Srikanth Madikeri, Teodora Vukovic, Volker Dellwo

The CL-UZH team submitted one system each for the fixed and open conditions of the NIST SRE 2024 challenge. For the closed-set condition, results for the audio-only trials were achieved using the X-vector system developed with Kaldi. For the audio-visual results we used only models developed for the visual modality. Two sets of results were submitted for the open-set and closed-set conditions, one based on a pretrained model using the VoxBlink2 and VoxCeleb2 datasets. An Xvector-based model was trained from scratch using the CTS superset dataset for the closed set. In addition to the submission of the results of the SRE24 evaluation to the competition website, we talked about the performance of the proposed systems on the SRE24 evaluation in this report.

comment: CL-UZH submission for the NIST SRE 2024 Evaluation plan

♻ ☆ Combining Deterministic Enhanced Conditions with Dual-Streaming Encoding for Diffusion-Based Speech Enhancement

Hao Shi, Xugang Lu, Kazuki Shimada, Tatsuya Kawahara

Diffusion-based speech enhancement (SE) models need to incorporate correct prior knowledge as reliable conditions to generate accurate predictions. However, providing reliable conditions using noisy features is challenging. One solution is to use features enhanced by deterministic methods as conditions. However, the information distortion and loss caused by deterministic methods might affect the diffusion process. In this paper, we first investigate the effects of using different deterministic SE models as conditions for diffusion. We validate two conditions depending on whether the noisy feature was used as part of the condition: one using only the deterministic feature (deterministic-only), and the other using both deterministic and noisy features (deterministic-noisy). Preliminary investigation found that using deterministic enhanced conditions improves hearing experiences on real data, while the choice between using deterministic-only or deterministic-noisy conditions depends on the deterministic models. Based on these findings, we propose a dual-streaming encoding Repair-Diffusion Model for SE (DERDM-SE) to more effectively utilize both conditions. Moreover, we found that fine-grained deterministic models have greater potential in objective evaluation metrics, while UNet-based deterministic models provide more stable diffusion performance. Therefore, in the DERDM-SE, we propose a deterministic model that combines coarse- and fine-grained processing. Experimental results on CHiME4 show that the proposed models effectively leverage deterministic models to achieve better SE evaluation scores, along with more stable performance compared to other diffusion-based SE models.

♻ ☆ An Investigation of Incorporating Mamba for Speech Enhancement SLT 2024

Rong Chao, Wen-Huang Cheng, Moreno La Quatra, Sabato Marco Siniscalchi, Chao-Han Huck Yang, Szu-Wei Fu, Yu Tsao

This work aims to investigate the use of a recently proposed, attention-free, scalable state-space model (SSM), Mamba, for the speech enhancement (SE) task. In particular, we employ Mamba to deploy different regression-based SE models (SEMamba) with different configurations, namely basic, advanced, causal, and non-causal. Furthermore, loss functions either based on signal-level distances or metric-oriented are considered. Experimental evidence shows that SEMamba attains a competitive PESQ of 3.55 on the VoiceBank-DEMAND dataset with the advanced, non-causal configuration. A new state-of-the-art PESQ of 3.69 is also reported when SEMamba is combined with Perceptual Contrast Stretching (PCS). Compared against Transformed-based equivalent SE solutions, a noticeable FLOPs reduction up to ~12% is observed with the advanced non-causal configurations. Finally, SEMamba can be used as a pre-processing step before automatic speech recognition (ASR), showing competitive performance against recent SE solutions.

comment: Accepted to IEEE SLT 2024

♻ ☆ Step-by-Step Video-to-Audio Synthesis via Negative Audio Guidance

Akio Hayakawa, Masato Ishii, Takashi Shibuya, Yuki Mitsufuji

We propose a step-by-step video-to-audio (V2A) generation method for finer controllability over the generation process and more realistic audio synthesis. Inspired by traditional Foley workflows, our approach aims to comprehensively capture all sound events induced by a video through the incremental generation of missing sound events. To avoid the need for costly multi-reference video-audio datasets, each generation step is formulated as a negatively guided V2A process that discourages duplication of existing sounds. The guidance model is trained by finetuning a pre-trained V2A model on audio pairs from adjacent segments of the same video, allowing training with standard single-reference audiovisual datasets that are easily accessible. Objective and subjective evaluations demonstrate that our method enhances the separability of generated sounds at each step and improves the overall quality of the final composite audio, outperforming existing baselines.

♻ ☆ Scattering Transformer: A Training-Free Transformer Architecture for Heart Murmur Detection

Rami Zewail

In an attempt to address the need for skilled clinicians in heart sound interpretation, recent research efforts on automating cardiac auscultation have explored deep learning approaches. The majority of these approaches have been based on supervised learning that is always challenged in occasions where training data is limited. More recently, there has been a growing interest in potentials of pre-trained self-supervised audio foundation models for biomedical end tasks. Despite exhibiting promising results, these foundational models are typically computationally intensive. Within the context of automatic cardiac auscultation, this study explores a lightweight alternative to these general-purpose audio foundation models by introducing the Scattering Transformer, a novel, training-free transformer architecture for heart murmur detection. The proposed method leverages standard wavelet scattering networks by introducing contextual dependencies in a transformer-like architecture without any backpropagation. We evaluate our approach on the public CirCor DigiScope dataset, directly comparing it against leading general-purpose foundational models. The Scattering Transformer achieves a Weighted Accuracy(WAR) of 0.786 and an Unweighted Average Recall(UAR) of 0.697, demonstrating performance highly competitive with contemporary state of the art methods. This study establishes the Scattering Transformer as a viable and promising alternative in resource-constrained setups.

comment: This paper has been accepted for presentation at the 14th International Conference on Model and Data Engineering (MEDI 2025). The final authenticated Version of Record will be published by Springer in the Lecture Notes in Computer Science (LNCS) series

♻ ☆ Large-Scale Training Data Attribution for Music Generative Models via Unlearning NeurIPS 2025

Woosung Choi, Junghyun Koo, Kin Wai Cheuk, Joan Serrà, Marco A. Martínez-Ramírez, Yukara Ikemiya, Naoki Murata, Yuhta Takida, Wei-Hsiang Liao, Yuki Mitsufuji

This paper explores the use of unlearning methods for training data attribution (TDA) in music generative models trained on large-scale datasets. TDA aims to identify which specific training data points contributed the most to the generation of a particular output from a specific model. This is crucial in the context of AI-generated music, where proper recognition and credit for original artists are generally overlooked. By enabling white-box attribution, our work supports a fairer system for acknowledging artistic contributions and addresses pressing concerns related to AI ethics and copyright. We apply unlearning-based attribution to a text-to-music diffusion model trained on a large-scale dataset and investigate its feasibility and behavior in this setting. To validate the method, we perform a grid search over different hyperparameter configurations and quantitatively evaluate the consistency of the unlearning approach. We then compare attribution patterns from unlearning with non-counterfactual approaches. Our findings suggest that unlearning-based approaches can be effectively adapted to music generative models, introducing large-scale TDA to this domain and paving the way for more ethical and accountable AI systems for music creation.

comment: accepted at NeurIPS 2025 Creative AI Track

♻ ☆ Synthetic Audio Forensics Evaluation (SAFE) Challenge

Kirill Trapeznikov, Paul Cummer, Pranay Pherwani, Jai Aslam, Michael S. Davinroy, Peter Bautista, Laura Cassani, Matthew Stamm, Jill Crisman

The increasing realism of synthetic speech generated by advanced text-to-speech (TTS) models, coupled with post-processing and laundering techniques, presents a significant challenge for audio forensic detection. In this paper, we introduce the SAFE (Synthetic Audio Forensics Evaluation) Challenge, a fully blind evaluation framework designed to benchmark detection models across progressively harder scenarios: raw synthetic speech, processed audio (e.g., compression, resampling), and laundered audio intended to evade forensic analysis. The SAFE challenge consisted of a total of 90 hours of audio and 21,000 audio samples split across 21 different real sources and 17 different TTS models and 3 tasks. We present the challenge, evaluation design and tasks, dataset details, and initial insights into the strengths and limitations of current approaches, offering a foundation for advancing synthetic audio detection research. More information is available at \href{https://stresearch.github.io/SAFE/}{https://stresearch.github.io/SAFE/}.

♻ ☆ Baseline Systems For The 2025 Low-Resource Audio Codec Challenge

Yusuf Ziya Isik, Rafał Łaganowski

The Low-Resource Audio Codec (LRAC) Challenge aims to advance neural audio coding for deployment in resource-constrained environments. The first edition focuses on low-resource neural speech codecs that must operate reliably under everyday noise and reverberation, while satisfying strict constraints on computational complexity, latency, and bitrate. Track 1 targets transparency codecs, which aim to preserve the perceptual transparency of input speech under mild noise and reverberation. Track 2 addresses enhancement codecs, which combine coding and compression with denoising and dereverberation. This paper presents the official baseline systems for both tracks in the 2025 LRAC Challenge. The baselines are convolutional neural codec models with Residual Vector Quantization, trained end-to-end using a combination of adversarial and reconstruction objectives. We detail the data filtering and augmentation strategies, model architectures, optimization procedures, and checkpoint selection criteria.

comment: Low-Resource Audio Codec Challenge 2025

☆ Modulation Discovery with Differentiable Digital Signal Processing SP

Christopher Mitcheltree, Hao Hao Tan, Joshua D. Reiss

Modulations are a critical part of sound design and music production, enabling the creation of complex and evolving audio. Modern synthesizers provide envelopes, low frequency oscillators (LFOs), and more parameter automation tools that allow users to modulate the output with ease. However, determining the modulation signals used to create a sound is difficult, and existing sound-matching / parameter estimation systems are often uninterpretable black boxes or predict high-dimensional framewise parameter values without considering the shape, structure, and routing of the underlying modulation curves. We propose a neural sound-matching approach that leverages modulation extraction, constrained control signal parameterizations, and differentiable digital signal processing (DDSP) to discover the modulations present in a sound. We demonstrate the effectiveness of our approach on highly modulated synthetic and real audio samples, its applicability to different DDSP synth architectures, and investigate the trade-off it incurs between interpretability and sound-matching accuracy. We make our code and audio samples available and provide the trained DDSP synths in a VST plugin.

comment: Accepted to WASPAA 2025 (best paper award candidate). Code, audio samples, and plugins can be found at https://christhetree.github.io/mod_discovery/

☆ TokenChain: A Discrete Speech Chain via Semantic Token Modeling ICASSP

Mingxuan Wang, Satoshi Nakamura

Machine Speech Chain, simulating the human perception-production loop, proves effective in jointly improving ASR and TTS. We propose TokenChain, a fully discrete speech chain coupling semantic-token ASR with a two-stage TTS: an autoregressive text-to-semantic model co-trained with ASR and a masked-generative semantic-to-acoustic model for synthesis only. End-to-end feedback across the text interface is enabled with straight-through argmax/Gumbel-Softmax and balanced with supervised ASR via dynamic weight averaging. Ablations examine optimal temperature schedules for in- and cross-domain transfer. Evaluation reveals TokenChain surpasses baseline accuracy 2-6 epochs earlier and yields 5-13% lower equal-epoch error with stable T2S on LibriSpeech, and reduces relative ASR WER by 56% and T2S WER by 31% on TED-LIUM with minimal forgetting, showing that chain learning remains effective with token interfaces and models.

comment: 5 pages, 3 figures. Submitted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026

☆ Latent Speech-Text Transformer

Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srinivasan Iyer, Duc Le

Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.

comment: 16 pages, 13 figures

☆ ECTSpeech: Enhancing Efficient Speech Synthesis via Easy Consistency Tuning

Tao Zhu, Yinfeng Yu, Liejun Wang, Fuchun Sun, Wendong Zheng

Diffusion models have demonstrated remarkable performance in speech synthesis, but typically require multi-step sampling, resulting in low inference efficiency. Recent studies address this issue by distilling diffusion models into consistency models, enabling efficient one-step generation. However, these approaches introduce additional training costs and rely heavily on the performance of pre-trained teacher models. In this paper, we propose ECTSpeech, a simple and effective one-step speech synthesis framework that, for the first time, incorporates the Easy Consistency Tuning (ECT) strategy into speech synthesis. By progressively tightening consistency constraints on a pre-trained diffusion model, ECTSpeech achieves high-quality one-step generation while significantly reducing training complexity. In addition, we design a multi-scale gate module (MSGate) to enhance the denoiser's ability to fuse features at different scales. Experimental results on the LJSpeech dataset demonstrate that ECTSpeech achieves audio quality comparable to state-of-the-art methods under single-step sampling, while substantially reducing the model's training cost and complexity.

comment: Accepted for publication by Proceedings of the 2025 ACM Multimedia Asia Conference(MMAsia '25)

☆ Revisiting Modeling and Evaluation Approaches in Speech Emotion Recognition: Considering Subjectivity of Annotators and Ambiguity of Emotions ACL

Huang-Cheng Chou, Chi-Chun Lee

Over the past two decades, speech emotion recognition (SER) has received growing attention. To train SER systems, researchers collect emotional speech databases annotated by crowdsourced or in-house raters who select emotions from predefined categories. However, disagreements among raters are common. Conventional methods treat these disagreements as noise, aggregating labels into a single consensus target. While this simplifies SER as a single-label task, it ignores the inherent subjectivity of human emotion perception. This dissertation challenges such assumptions and asks: (1) Should minority emotional ratings be discarded? (2) Should SER systems learn from only a few individuals' perceptions? (3) Should SER systems predict only one emotion per sample? Psychological studies show that emotion perception is subjective and ambiguous, with overlapping emotional boundaries. We propose new modeling and evaluation perspectives: (1) Retain all emotional ratings and represent them with soft-label distributions. Models trained on individual annotator ratings and jointly optimized with standard SER systems improve performance on consensus-labeled tests. (2) Redefine SER evaluation by including all emotional data and allowing co-occurring emotions (e.g., sad and angry). We propose an ``all-inclusive rule'' that aggregates all ratings to maximize diversity in label representation. Experiments on four English emotion databases show superior performance over majority and plurality labeling. (3) Construct a penalization matrix to discourage unlikely emotion combinations during training. Integrating it into loss functions further improves performance. Overall, embracing minority ratings, multiple annotators, and multi-emotion predictions yields more robust and human-aligned SER systems.

comment: PhD Thesis; ACLCLP Doctoral Dissertation Award -- Honorable Mention

☆ Revisiting MFCCs: Evidence for Spectral-Prosodic Coupling SC

Vitor Magno de O. S. Bezerra, Gabriel F. A. Bastos, Jugurta Montalvão

Mel-frequency cepstral coefficients (MFCCs) are an important feature in speech processing. A deeper understanding of their properties can contribute to the work that is being done with both classical and deep learning models. This study challenges the long-held assumption that MFCCs lack relevant temporal information by investigating their relationship with speech prosody. Using a null hypothesis significance testing framework, a systematic assessment is made about the statistical independence between MFCCs and the three prosodic features: energy, fundamental frequency (F0), and voicing. The results demonstrate that it is statistically implausible that the MFCCs are independent of any of these three prosodic features. This finding suggests that MFCCs inherently carry valuable prosodic information, which can inform the design of future models in speech analysis and recognition.

comment: 5 pages, 3 figures, ISCMI 2025

☆ Segment-Factorized Full-Song Generation on Symbolic Piano Music NeurIPS 2025

Ping-Yi Chen, Chih-Pin Tan, Yi-Hsuan Yang

We propose the Segmented Full-Song Model (SFS) for symbolic full-song generation. The model accepts a user-provided song structure and an optional short seed segment that anchors the main idea around which the song is developed. By factorizing a song into segments and generating each one through selective attention to related segments, the model achieves higher quality and efficiency compared to prior work. To demonstrate its suitability for human-AI interaction, we further wrap SFS into a web application that enables users to iteratively co-create music on a piano roll with customizable structures and flexible ordering.

comment: Accepted to the 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: AI for Music

☆ FoleyGRAM: Video-to-Audio Generation with GRAM-Aligned Multimodal Encoders IJCNN 2025

Riccardo Fosco Gramaccioni, Christian Marinoni, Eleonora Grassucci, Giordano Cicchetti, Aurelio Uncini, Danilo Comminiello

In this work, we present FoleyGRAM, a novel approach to video-to-audio generation that emphasizes semantic conditioning through the use of aligned multimodal encoders. Building on prior advancements in video-to-audio generation, FoleyGRAM leverages the Gramian Representation Alignment Measure (GRAM) to align embeddings across video, text, and audio modalities, enabling precise semantic control over the audio generation process. The core of FoleyGRAM is a diffusion-based audio synthesis model conditioned on GRAM-aligned embeddings and waveform envelopes, ensuring both semantic richness and temporal alignment with the corresponding input video. We evaluate FoleyGRAM on the Greatest Hits dataset, a standard benchmark for video-to-audio models. Our experiments demonstrate that aligning multimodal encoders using GRAM enhances the system's ability to semantically align generated audio with video content, advancing the state of the art in video-to-audio synthesis.

comment: Acepted at IJCNN 2025

☆ StereoSync: Spatially-Aware Stereo Audio Generation from Video IJCNN 2025

Christian Marinoni, Riccardo Fosco Gramaccioni, Kazuki Shimada, Takashi Shibuya, Yuki Mitsufuji, Danilo Comminiello

Although audio generation has been widely studied over recent years, video-aligned audio generation still remains a relatively unexplored frontier. To address this gap, we introduce StereoSync, a novel and efficient model designed to generate audio that is both temporally synchronized with a reference video and spatially aligned with its visual context. Moreover, StereoSync also achieves efficiency by leveraging pretrained foundation models, reducing the need for extensive training while maintaining high-quality synthesis. Unlike existing methods that primarily focus on temporal synchronization, StereoSync introduces a significant advancement by incorporating spatial awareness into video-aligned audio generation. Indeed, given an input video, our approach extracts spatial cues from depth maps and bounding boxes, using them as cross-attention conditioning in a diffusion-based audio generation model. Such an approach allows StereoSync to go beyond simple synchronization, producing stereo audio that dynamically adapts to the spatial structure and movement of a video scene. We evaluate StereoSync on Walking The Maps, a curated dataset comprising videos from video games that feature animated characters walking through diverse environments. Experimental results demonstrate the ability of StereoSync to achieve both temporal and spatial alignment, advancing the state of the art in video-to-audio generation and resulting in a significantly more immersive and realistic audio experience.

comment: Accepted at IJCNN 2025

☆ Neural Forward Filtering for Speaker-Image Separation

Jingqi Sun, Shulin He, Ruizhe Pang, Zhong-Qiu Wang

We address monaural multi-speaker-image separation in reverberant conditions, aiming at separating mixed speakers but preserving the reverberation of each speaker. A straightforward approach for this task is to directly train end-to-end DNN systems to predict the reverberant speech of each speaker based on the input mixture. Although effective, this approach does not explicitly exploit the physical constraint that reverberant speech can be reproduced by convolving the direct-path signal with a linear filter. To address this, we propose CxNet, a two-DNN system with a neural forward filtering module in between. The first DNN is trained to jointly predict the direct-path signal and reverberant speech. Based on the direct-path estimate, the neural forward filtering module estimates the linear filter, and the estimated filter is then convolved with the direct-path estimate to obtain another estimate of reverberant speech, which is utilized as a discriminative feature to help the second DNN better estimate the reverberant speech. By explicitly modeling the linear filter, CxNet could leverage the physical constraint between the direct-path signal and reverberant speech to capture crucial information about reverberation tails. Evaluation results on the SMS-WSJ dataset show the effectiveness of the proposed algorithms.

comment: in submission

☆ Transcribing Rhythmic Patterns of the Guitar Track in Polyphonic Music SP

Aleksandr Lukoianov, Anssi Klapuri

Whereas chord transcription has received considerable attention during the past couple of decades, far less work has been devoted to transcribing and encoding the rhythmic patterns that occur in a song. The topic is especially relevant for instruments such as the rhythm guitar, which is typically played by strumming rhythmic patterns that repeat and vary over time. However, in many cases one cannot objectively define a single "right" rhythmic pattern for a given song section. To create a dataset with well-defined ground-truth labels, we asked expert musicians to transcribe the rhythmic patterns in 410 popular songs and record cover versions where the guitar tracks followed those transcriptions. To transcribe the strums and their corresponding rhythmic patterns, we propose a three-step framework. Firstly, we perform approximate stem separation to extract the guitar part from the polyphonic mixture. Secondly, we detect individual strums within the separated guitar audio, using a pre-trained foundation model (MERT) as a backbone. Finally, we carry out a pattern-decoding process in which the transcribed sequence of guitar strums is represented by patterns drawn from an expert-curated vocabulary. We show that it is possible to transcribe the rhythmic patterns of the guitar track in polyphonic music with quite high accuracy, producing a representation that is human-readable and includes automatically detected bar lines and time signature markers. We perform ablation studies and error analysis and propose a set of evaluation metrics to assess the accuracy and readability of the predicted rhythmic pattern sequence.

comment: Accepted to WASPAA 2025

☆ Investigation of perception inconsistency in speaker embedding for asynchronous voice anonymization

Rui Wang, Liping Chen, Kong Aik Lee, Zhengpeng Zha, Zhenhua Ling

Given the speech generation framework that represents the speaker attribute with an embedding vector, asynchronous voice anonymization can be achieved by modifying the speaker embedding derived from the original speech. However, the inconsistency between machine and human perceptions of the speaker attribute within the speaker embedding remains unexplored, limiting its performance in asynchronous voice anonymization. To this end, this study investigates this inconsistency via modifications to speaker embedding in the speech generation process. Experiments conducted on the FACodec and Diff-HierVC speech generation models discover a subspace whose removal alters machine perception while preserving its human perception of the speaker attribute in the generated speech. With these findings, an asynchronous voice anonymization is developed, achieving 100% human perception preservation rate while obscuring the machine perception. Audio samples can be found in https://voiceprivacy.github.io/speaker-embedding-eigen-decomposition/.

☆ Teaching Machines to Speak Using Articulatory Control

Akshay Anand, Chenxu Guo, Cheol Jun Cho, Jiachen Lian, Gopala Anumanchipalli

Current speech production systems predominantly rely on large transformer models that operate as black boxes, providing little interpretability or grounding in the physical mechanisms of human speech. We address this limitation by proposing a new framework: speech generation through explicit articulatory control. This reframes speech as a motor control task similar to robotic manipulation. Our approach uses reinforcement learning to train a policy that directly controls the movements of vocal tract articulators, such as the tongue, lips, and jaw, to produce syllable-level speech. Specifically, we employ the Proximal Policy Optimization algorithm to learn optimal articulatory movements based on acoustic feedback provided by our audio perceiver, Sylber. The resulting articulatory trajectories are decoded into audio using SPARC, a pre-trained articulatory-to-speech decoder. We train this framework on six target syllables, and it demonstrates successful convergence, with similarity scores between the policy-generated audio and the target syllables exceeding 0.85. Accurate human transcription of the audio for syllables such as "please", "loot", and "cat" demonstrates the intelligibility of this framework.

☆ Sci-Phi: A Large Language Model Spatial Audio Descriptor

Xilin Jiang, Hannes Gamper, Sebastian Braun

Acoustic scene perception involves describing the type of sounds, their timing, their direction and distance, as well as their loudness and reverberation. While audio language models excel in sound recognition, single-channel input fundamentally limits spatial understanding. This work presents Sci-Phi, a spatial audio large language model with dual spatial and spectral encoders that estimates a complete parameter set for all sound sources and the surrounding environment. Learning from over 4,000 hours of synthetic first-order Ambisonics recordings including metadata, Sci-Phi enumerates and describes up to four directional sound sources in one pass, alongside non-directional background sounds and room characteristics. We evaluate the model with a permutation-invariant protocol and 15 metrics covering content, location, timing, loudness, and reverberation, and analyze its robustness across source counts, signal-to-noise ratios, reverberation levels, and challenging mixtures of acoustically, spatially, or temporally similar sources. Notably, Sci-Phi generalizes to real room impulse responses with only minor performance degradation. Overall, this work establishes the first audio LLM capable of full spatial-scene description, with strong potential for real-world deployment. Demo: https://sci-phi-audio.github.io/demo

☆ AQA-TTRL: Self-Adaptation in Audio Question Answering with Test-Time Reinforcement Learning ICASSP 2026

Haoyu Zhang, Jiaxian Guo, Yusuke Iwasawa, Yutaka Matsuo

Large Audio Language Models (LALMs) demonstrate impressive general audio understanding, but once deployed, they are static and fail to improve with new real-world audio data. As traditional supervised fine-tuning is costly, we introduce a novel framework for test-time audio understanding, AQA-TTRL, where an LALM evolves on-the-fly using only unlabeled test data. It first generates pseudo-labels from the prediction via majority voting, then optimizes the model via reinforcement learning. To handle the inherent noise in these self-generated labels, we introduce a confidence-based weighting method to adjust training signals. Furthermore, a multiple-attempt sampling operation mitigates advantage collapse and stabilizes training. On the MMAU (test-mini/test), MMAR, and MMSU benchmarks, AQA-TTRL achieves significant average improvements of 4.42% for the Qwen2.5-Omni 7B model and 11.04% for the 3B model. Notably, the adapted 3B model consistently outperforms the direct inference of the unadapted 7B model, highlighting the effectiveness of previously unexplored test-time adaptations in audio understanding.

comment: 5 pages, 4 figures, Submitted to ICASSP 2026

♻ ☆ MuFFIN: Multifaceted Pronunciation Feedback Model with Interactive Hierarchical Neural Modeling

Bi-Cheng Yan, Ming-Kang Tsai, Berlin Chen

Computer-assisted pronunciation training (CAPT) manages to facilitate second-language (L2) learners to practice pronunciation skills by offering timely and instructive feedback. To examine pronunciation proficiency from multiple facets, existing methods for CAPT broadly fall into two categories: mispronunciation detection and diagnosis (MDD) as well as automatic pronunciation assessment (APA). The former aims to pinpoint phonetic pronunciation errors and provide diagnostic feedback, while the latter seeks instead to quantify pronunciation proficiency pertaining to various aspects. Despite the natural complementarity between MDD and APA, researchers and practitioners, however, often treat them as independent tasks with disparate modeling paradigms. In light of this, we in this paper first introduce MuFFIN, a Multi-Faceted pronunciation Feedback model with an Interactive hierarchical Neural architecture, to jointly address the tasks of MDD and APA. To better capture the nuanced distinctions between phonemes in the feature space, a novel phoneme-contrastive ordinal regularization mechanism is then put forward to optimize the proposed model to generate more phoneme-discriminative features while factoring in the ordinality of the aspect scores. In addition, to address the intricate data imbalance problem in MDD, we design a simple yet effective training objective, which is specifically tailored to perturb the outputs of a phoneme classifier with the phoneme-specific variations, so as to better render the distribution of predicted phonemes meanwhile considering their mispronunciation characteristics. A series of experiments conducted on the Speechocean762 benchmark dataset demonstrates the efficacy of our method in relation to several cutting-edge baselines, showing state-of-the-art performance on both the APA and MDD tasks.

comment: Accepted and to appear in IEEE/ACM Transactions on Audio, Speech, and Language Processing

♻ ☆ CL-UZH submission to the NIST SRE 2024 Speaker Recognition Evaluation

Aref Farhadipour, Shiran Liu, Masoumeh Chapariniya, Valeriia Vyshnevetska, Srikanth Madikeri, Teodora Vukovic, Volker Dellwo

The CL-UZH team submitted one system each for the fixed and open conditions of the NIST SRE 2024 challenge. For the closed-set condition, results for the audio-only trials were achieved using the X-vector system developed with Kaldi. For the audio-visual results we used only models developed for the visual modality. Two sets of results were submitted for the open-set and closed-set conditions, one based on a pretrained model using the VoxBlink2 and VoxCeleb2 datasets. An Xvector-based model was trained from scratch using the CTS superset dataset for the closed set. In addition to the submission of the results of the SRE24 evaluation to the competition website, we talked about the performance of the proposed systems on the SRE24 evaluation in this report.

comment: CL-UZH submission for the NIST SRE 2024 Evaluation plan

♻ ☆ Combining Deterministic Enhanced Conditions with Dual-Streaming Encoding for Diffusion-Based Speech Enhancement

Hao Shi, Xugang Lu, Kazuki Shimada, Tatsuya Kawahara

Diffusion-based speech enhancement (SE) models need to incorporate correct prior knowledge as reliable conditions to generate accurate predictions. However, providing reliable conditions using noisy features is challenging. One solution is to use features enhanced by deterministic methods as conditions. However, the information distortion and loss caused by deterministic methods might affect the diffusion process. In this paper, we first investigate the effects of using different deterministic SE models as conditions for diffusion. We validate two conditions depending on whether the noisy feature was used as part of the condition: one using only the deterministic feature (deterministic-only), and the other using both deterministic and noisy features (deterministic-noisy). Preliminary investigation found that using deterministic enhanced conditions improves hearing experiences on real data, while the choice between using deterministic-only or deterministic-noisy conditions depends on the deterministic models. Based on these findings, we propose a dual-streaming encoding Repair-Diffusion Model for SE (DERDM-SE) to more effectively utilize both conditions. Moreover, we found that fine-grained deterministic models have greater potential in objective evaluation metrics, while UNet-based deterministic models provide more stable diffusion performance. Therefore, in the DERDM-SE, we propose a deterministic model that combines coarse- and fine-grained processing. Experimental results on CHiME4 show that the proposed models effectively leverage deterministic models to achieve better SE evaluation scores, along with more stable performance compared to other diffusion-based SE models.

♻ ☆ An Investigation of Incorporating Mamba for Speech Enhancement SLT 2024

Rong Chao, Wen-Huang Cheng, Moreno La Quatra, Sabato Marco Siniscalchi, Chao-Han Huck Yang, Szu-Wei Fu, Yu Tsao

This work aims to investigate the use of a recently proposed, attention-free, scalable state-space model (SSM), Mamba, for the speech enhancement (SE) task. In particular, we employ Mamba to deploy different regression-based SE models (SEMamba) with different configurations, namely basic, advanced, causal, and non-causal. Furthermore, loss functions either based on signal-level distances or metric-oriented are considered. Experimental evidence shows that SEMamba attains a competitive PESQ of 3.55 on the VoiceBank-DEMAND dataset with the advanced, non-causal configuration. A new state-of-the-art PESQ of 3.69 is also reported when SEMamba is combined with Perceptual Contrast Stretching (PCS). Compared against Transformed-based equivalent SE solutions, a noticeable FLOPs reduction up to ~12% is observed with the advanced non-causal configurations. Finally, SEMamba can be used as a pre-processing step before automatic speech recognition (ASR), showing competitive performance against recent SE solutions.

comment: Accepted to IEEE SLT 2024

♻ ☆ Speech-Based Cognitive Screening: A Systematic Evaluation of LLM Adaptation Strategies

Fatemeh Taherinezhad, Mohamad Javad Momeni Nezhad, Sepehr Karimi, Sina Rashidi, Ali Zolnour, Maryam Dadkhah, Yasaman Haghbin, Hossein AzadMaleki, Maryam Zolnoori

Over half of US adults with Alzheimer disease and related dementias remain undiagnosed, and speech-based screening offers a scalable detection approach. We compared large language model adaptation strategies for dementia detection using the DementiaBank speech corpus, evaluating nine text-only models and three multimodal audio-text models on recordings from DementiaBank speech corpus. Adaptations included in-context learning with different demonstration selection policies, reasoning-augmented prompting, parameter-efficient fine-tuning, and multimodal integration. Results showed that class-centroid demonstrations achieved the highest in-context learning performance, reasoning improved smaller models, and token-level fine-tuning generally produced the best scores. Adding a classification head substantially improved underperforming models. Among multimodal models, fine-tuned audio-text systems performed well but did not surpass the top text-only models. These findings highlight that model adaptation strategies, including demonstration selection, reasoning design, and tuning method, critically influence speech-based dementia detection, and that properly adapted open-weight models can match or exceed commercial systems.

♻ ☆ Step-by-Step Video-to-Audio Synthesis via Negative Audio Guidance

Akio Hayakawa, Masato Ishii, Takashi Shibuya, Yuki Mitsufuji

We propose a step-by-step video-to-audio (V2A) generation method for finer controllability over the generation process and more realistic audio synthesis. Inspired by traditional Foley workflows, our approach aims to comprehensively capture all sound events induced by a video through the incremental generation of missing sound events. To avoid the need for costly multi-reference video-audio datasets, each generation step is formulated as a negatively guided V2A process that discourages duplication of existing sounds. The guidance model is trained by finetuning a pre-trained V2A model on audio pairs from adjacent segments of the same video, allowing training with standard single-reference audiovisual datasets that are easily accessible. Objective and subjective evaluations demonstrate that our method enhances the separability of generated sounds at each step and improves the overall quality of the final composite audio, outperforming existing baselines.

♻ ☆ Scattering Transformer: A Training-Free Transformer Architecture for Heart Murmur Detection

Rami Zewail

In an attempt to address the need for skilled clinicians in heart sound interpretation, recent research efforts on automating cardiac auscultation have explored deep learning approaches. The majority of these approaches have been based on supervised learning that is always challenged in occasions where training data is limited. More recently, there has been a growing interest in potentials of pre-trained self-supervised audio foundation models for biomedical end tasks. Despite exhibiting promising results, these foundational models are typically computationally intensive. Within the context of automatic cardiac auscultation, this study explores a lightweight alternative to these general-purpose audio foundation models by introducing the Scattering Transformer, a novel, training-free transformer architecture for heart murmur detection. The proposed method leverages standard wavelet scattering networks by introducing contextual dependencies in a transformer-like architecture without any backpropagation. We evaluate our approach on the public CirCor DigiScope dataset, directly comparing it against leading general-purpose foundational models. The Scattering Transformer achieves a Weighted Accuracy(WAR) of 0.786 and an Unweighted Average Recall(UAR) of 0.697, demonstrating performance highly competitive with contemporary state of the art methods. This study establishes the Scattering Transformer as a viable and promising alternative in resource-constrained setups.

comment: This paper has been accepted for presentation at the 14th International Conference on Model and Data Engineering (MEDI 2025). The final authenticated Version of Record will be published by Springer in the Lecture Notes in Computer Science (LNCS) series

♻ ☆ Large-Scale Training Data Attribution for Music Generative Models via Unlearning NeurIPS 2025

Woosung Choi, Junghyun Koo, Kin Wai Cheuk, Joan Serrà, Marco A. Martínez-Ramírez, Yukara Ikemiya, Naoki Murata, Yuhta Takida, Wei-Hsiang Liao, Yuki Mitsufuji

This paper explores the use of unlearning methods for training data attribution (TDA) in music generative models trained on large-scale datasets. TDA aims to identify which specific training data points contributed the most to the generation of a particular output from a specific model. This is crucial in the context of AI-generated music, where proper recognition and credit for original artists are generally overlooked. By enabling white-box attribution, our work supports a fairer system for acknowledging artistic contributions and addresses pressing concerns related to AI ethics and copyright. We apply unlearning-based attribution to a text-to-music diffusion model trained on a large-scale dataset and investigate its feasibility and behavior in this setting. To validate the method, we perform a grid search over different hyperparameter configurations and quantitatively evaluate the consistency of the unlearning approach. We then compare attribution patterns from unlearning with non-counterfactual approaches. Our findings suggest that unlearning-based approaches can be effectively adapted to music generative models, introducing large-scale TDA to this domain and paving the way for more ethical and accountable AI systems for music creation.

comment: accepted at NeurIPS 2025 Creative AI Track

♻ ☆ Synthetic Audio Forensics Evaluation (SAFE) Challenge

Kirill Trapeznikov, Paul Cummer, Pranay Pherwani, Jai Aslam, Michael S. Davinroy, Peter Bautista, Laura Cassani, Matthew Stamm, Jill Crisman

The increasing realism of synthetic speech generated by advanced text-to-speech (TTS) models, coupled with post-processing and laundering techniques, presents a significant challenge for audio forensic detection. In this paper, we introduce the SAFE (Synthetic Audio Forensics Evaluation) Challenge, a fully blind evaluation framework designed to benchmark detection models across progressively harder scenarios: raw synthetic speech, processed audio (e.g., compression, resampling), and laundered audio intended to evade forensic analysis. The SAFE challenge consisted of a total of 90 hours of audio and 21,000 audio samples split across 21 different real sources and 17 different TTS models and 3 tasks. We present the challenge, evaluation design and tasks, dataset details, and initial insights into the strengths and limitations of current approaches, offering a foundation for advancing synthetic audio detection research. More information is available at \href{https://stresearch.github.io/SAFE/}{https://stresearch.github.io/SAFE/}.

☆ Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba

Baher Mohammad, Magauiya Zhussip, Stamatios Lefkimmiatis

We introduce MAVE (Mamba with Cross-Attention for Voice Editing and Synthesis), a novel autoregressive architecture for text-conditioned voice editing and high-fidelity text-to-speech (TTS) synthesis, built on a cross-attentive Mamba backbone. MAVE achieves state-of-the-art performance in speech editing and very competitive results in zero-shot TTS, while not being explicitly trained on the latter task, outperforming leading autoregressive and diffusion models on diverse, real-world audio. By integrating Mamba for efficient audio sequence modeling with cross-attention for precise text-acoustic alignment, MAVE enables context-aware voice editing with exceptional naturalness and speaker consistency. In pairwise human evaluations on a random 40-sample subset of the RealEdit benchmark (400 judgments), 57.2% of listeners rated MAVE - edited speech as perceptually equal to the original, while 24.8% prefered the original and 18.0% MAVE - demonstrating that in the majority of cases edits are indistinguishable from the source. MAVE compares favorably with VoiceCraft and FluentSpeech both on pairwise comparisons and standalone mean opinion score (MOS) evaluations. For zero-shot TTS, MAVE exceeds VoiceCraft in both speaker similarity and naturalness, without requiring multiple inference runs or post-processing. Remarkably, these quality gains come with a significantly lower memory cost and approximately the same latency: MAVE requires ~6x less memory than VoiceCraft during inference on utterances from the RealEdit database (mean duration: 6.21s, A100, FP16, batch size 1). Our results demonstrate that MAVE establishes a new standard for flexible, high-fidelity voice editing and synthesis through the synergistic integration of structured state-space modeling and cross-modal attention.

☆ A Study on the Data Distribution Gap in Music Emotion Recognition

Joann Ching, Gerhard Widmer

Music Emotion Recognition (MER) is a task deeply connected to human perception, relying heavily on subjective annotations collected from contributors. Prior studies tend to focus on specific musical styles rather than incorporating a diverse range of genres, such as rock and classical, within a single framework. In this paper, we address the task of recognizing emotion from audio content by investigating five datasets with dimensional emotion annotations -- EmoMusic, DEAM, PMEmo, WTC, and WCMED -- which span various musical styles. We demonstrate the problem of out-of-distribution generalization in a systematic experiment. By closely looking at multiple data and feature sets, we provide insight into genre-emotion relationships in existing data and examine potential genre dominance and dataset biases in certain feature representations. Based on these experiments, we arrive at a simple yet effective framework that combines embeddings extracted from the Jukebox model with chroma features and demonstrate how, alongside a combination of several diverse training sets, this permits us to train models with substantially improved cross-dataset generalization capabilities.

comment: Accepted at the 17th International Symposium on Computer Music Multidisciplinary Research (CMMR) 2025

☆ UniVoice: Unifying Autoregressive ASR and Flow-Matching based TTS with Large Language Models

Wenhao Guan, Zhikang Niu, Ziyue Jiang, Kaidi Wang, Peijie Chen, Qingyang Hong, Lin Li, Xie Chen

Large language models (LLMs) have demonstrated promising performance in both automatic speech recognition (ASR) and text-to-speech (TTS) systems, gradually becoming the mainstream approach. However, most current approaches address these tasks separately rather than through a unified framework. This work aims to integrate these two tasks into one unified model. Although discrete speech tokenization enables joint modeling, its inherent information loss limits performance in both recognition and generation. In this work, we present UniVoice, a unified LLM framework through continuous representations that seamlessly integrates speech recognition and synthesis within a single model. Our approach combines the strengths of autoregressive modeling for speech recognition with flow matching for high-quality generation. To mitigate the inherent divergence between autoregressive and flow-matching models, we further design a dual attention mechanism, which switches between a causal mask for recognition and a bidirectional attention mask for synthesis. Furthermore, the proposed text-prefix-conditioned speech infilling method enables high-fidelity zero-shot voice cloning. Experimental results demonstrate that our method can achieve or exceed current single-task modeling methods in both ASR and zero-shot TTS tasks. This work explores new possibilities for end-to-end speech understanding and generation.

☆ Robustness assessment of large audio language models in multiple-choice evaluation ICASSP 2026

Fernando López, Santosh Kesiraju, Jordi Luque

Recent advances in large audio language models (LALMs) have primarily been assessed using a multiple-choice question answering (MCQA) framework. However, subtle changes, such as shifting the order of choices, result in substantially different results. Existing MCQA frameworks do not account for this variability and report a single accuracy number per benchmark or category. We dive into the MCQA evaluation framework and conduct a systematic study spanning three benchmarks (MMAU, MMAR and MMSU) and four models: Audio Flamingo 2, Audio Flamingo 3, Qwen2.5-Omni-7B-Instruct, and Kimi-Audio-7B-Instruct. Our findings indicate that models are sensitive not only to the ordering of choices, but also to the paraphrasing of the question and the choices. Finally, we propose a simpler evaluation protocol and metric that account for subtle variations and provide a more detailed evaluation report of LALMs within the MCQA framework.

comment: Submitted to ICASSP 2026

☆ Language Model Based Text-to-Audio Generation: Anti-Causally Aligned Collaborative Residual Transformers EMNLP 2025

Juncheng Wang, Chao Xu, Cheng Yu, Zhe Hu, Haoyu Xie, Guoqi Yu, Lei Shang, Shujun Wang

While language models (LMs) paired with residual vector quantization (RVQ) tokenizers have shown promise in text-to-audio (T2A) generation, they still lag behind diffusion-based models by a non-trivial margin. We identify a critical dilemma underpinning this gap: incorporating more RVQ layers improves audio reconstruction fidelity but exceeds the generation capacity of conventional LMs. To address this, we first analyze RVQ dynamics and uncover two key limitations: 1) orthogonality of features across RVQ layers hinders effective LMs training, and 2) descending semantic richness in tokens from deeper RVQ layers exacerbates exposure bias during autoregressive decoding. Based on these insights, we propose Siren, a novel LM-based framework that employs multiple isolated transformers with causal conditioning and anti-causal alignment via reinforcement learning. Extensive experiments demonstrate that Siren outperforms both existing LM-based and diffusion-based T2A systems, achieving state-of-the-art results. By bridging the representational strengths of LMs with the fidelity demands of audio synthesis, our approach repositions LMs as competitive contenders against diffusion models in T2A tasks. Moreover, by aligning audio representations with linguistic structures, Siren facilitates a promising pathway toward unified multi-modal generation frameworks.

comment: Accepted to EMNLP 2025

☆ Evaluating Self-Supervised Speech Models via Text-Based LLMS ASRU 2025

Takashi Maekaku, Keita Goto, Jinchuan Tian, Yusuke Shinohara, Shinji Watanabe

Self-Supervised Learning (SSL) has gained traction for its ability to learn rich representations with low labeling costs, applicable across diverse downstream tasks. However, assessing the downstream-task performance remains challenging due to the cost of extra training and evaluation. Existing methods for task-agnostic evaluation also require extra training or hyperparameter tuning. We propose a novel evaluation metric using large language models (LLMs). By inputting discrete token sequences and minimal domain cues derived from SSL models into LLMs, we obtain the mean log-likelihood; these cues guide in-context learning, rendering the score more reliable without extra training or hyperparameter tuning. Experimental results show a correlation between LLM-based scores and automatic speech recognition task. Additionally, our findings reveal that LLMs not only functions as an SSL evaluation tools but also provides inference-time embeddings that are useful for speaker verification task.

comment: Accepted to ASRU 2025

☆ Differentiable physics for sound field reconstruction

Samuel A. Verburg, Efren Fernandez-Grande, Peter Gerstoft

Sound field reconstruction involves estimating sound fields from a limited number of spatially distributed observations. This work introduces a differentiable physics approach for sound field reconstruction, where the initial conditions of the wave equation are approximated with a neural network, and the differential operator is computed with a differentiable numerical solver. The use of a numerical solver enables a stable network training while enforcing the physics as a strong constraint, in contrast to conventional physics-informed neural networks, which include the physics as a constraint in the loss function. We introduce an additional sparsity-promoting constraint to achieve meaningful solutions even under severe undersampling conditions. Experiments demonstrate that the proposed approach can reconstruct sound fields under extreme data scarcity, achieving higher accuracy and better convergence compared to physics-informed neural networks.

comment: 28 pages plus references, 8 figures, full journal paper

☆ AUREXA-SE: Audio-Visual Unified Representation Exchange Architecture with Cross-Attention and Squeezeformer for Speech Enhancement

M. Sajid, Deepanshu Gupta, Yash Modi, Sanskriti Jain, Harshith Jai Surya Ganji, A. Rahaman, Harshvardhan Choudhary, Nasir Saleem, Amir Hussain, M. Tanveer

In this paper, we propose AUREXA-SE (Audio-Visual Unified Representation Exchange Architecture with Cross-Attention and Squeezeformer for Speech Enhancement), a progressive bimodal framework tailored for audio-visual speech enhancement (AVSE). AUREXA-SE jointly leverages raw audio waveforms and visual cues by employing a U-Net-based 1D convolutional encoder for audio and a Swin Transformer V2 for efficient and expressive visual feature extraction. Central to the architecture is a novel bidirectional cross-attention mechanism, which facilitates deep contextual fusion between modalities, enabling rich and complementary representation learning. To capture temporal dependencies within the fused embeddings, a stack of lightweight Squeezeformer blocks combining convolutional and attention modules is introduced. The enhanced embeddings are then decoded via a U-Net-style decoder for direct waveform reconstruction, ensuring perceptually consistent and intelligible speech output. Experimental evaluations demonstrate the effectiveness of AUREXA-SE, achieving significant performance improvements over noisy baselines, with STOI of 0.516, PESQ of 1.323, and SI-SDR of -4.322 dB. The source code of AUREXA-SE is available at https://github.com/mtanveer1/AVSEC-4-Challenge-2025.

☆ Provable Speech Attributes Conversion via Latent Independence

Jonathan Svirsky, Ofir Lindenbaum, Uri Shaham

While signal conversion and disentangled representation learning have shown promise for manipulating data attributes across domains such as audio, image, and multimodal generation, existing approaches, especially for speech style conversion, are largely empirical and lack rigorous theoretical foundations to guarantee reliable and interpretable control. In this work, we propose a general framework for speech attribute conversion, accompanied by theoretical analysis and guarantees under reasonable assumptions. Our framework builds on a non-probabilistic autoencoder architecture with an independence constraint between the predicted latent variable and the target controllable variable. This design ensures a consistent signal transformation, conditioned on an observed style variable, while preserving the original content and modifying the desired attribute. We further demonstrate the versatility of our method by evaluating it on speech styles, including speaker identity and emotion. Quantitative evaluations confirm the effectiveness and generality of the proposed approach.

☆ BaldWhisper: Faster Whisper with Head Shearing and Layer Merging

Yaya Sy, Christophe Cerisara, Irina Illina

Pruning large pre-trained transformers for low-resource languages is challenging, as it often requires massive retraining data to recover performance. For instance, Distill-Whisper prunes Whisper by 40% and retrains on 21,000 hours of speech, far beyond what is available for most languages. Can Whisper be made lighter and faster for edge devices in data-scarce settings? Focusing on Bambara with only 32h of speech-to-text data, we propose a new pruning recipe. Instead of vocabulary pruning, which is unsuitable due to frequent code-switching by Bambara speakers, we compress the embeddings with low-rank decomposition and feature distillation. Rather than removing layers, we merge them to limit performance loss. The final model preserves 90% of the original performance while being 48% smaller and 2.15x faster on a MacBook Air M1.

♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS

Seungheon Doh, Keunwoo Choi, Juhan Nam

While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.

comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)

♻ ☆ Latent Multi-view Learning for Robust Environmental Sound Representations DCASE 2025

Sivan Ding, Julia Wilkins, Magdalena Fuentes, Juan Pablo Bello

Self-supervised learning (SSL) approaches, such as contrastive and generative methods, have advanced environmental sound representation learning using unlabeled data. However, how these approaches can complement each other within a unified framework remains relatively underexplored. In this work, we propose a multi-view learning framework that integrates contrastive principles into a generative pipeline to capture sound source and device information. Our method encodes compressed audio latents into view-specific and view-common subspaces, guided by two self-supervised objectives: contrastive learning for targeted information flow between subspaces, and reconstruction for overall information preservation. We evaluate our method on an urban sound sensor network dataset for sound source and sensor classification, demonstrating improved downstream performance over traditional SSL techniques. Additionally, we investigate the model's potential to disentangle environmental sound attributes within the structured latent space under varied training configurations.

comment: Accepted to DCASE 2025 Workshop. 4+1 pages, 2 figures, 2 tables

♻ ☆ Low Resource Audio Codec Challenge Baseline Systems

Yusuf Ziya Isik, Rafał Łaganowski

The Low-Resource Audio Codec (LRAC) Challenge aims to advance neural audio coding for deployment in resource-constrained environments. The first edition focuses on low-resource neural speech codecs that must operate reliably under everyday noise and reverberation, while satisfying strict constraints on computational complexity, latency, and bitrate. Track 1 targets transparency codecs, which aim to preserve the perceptual transparency of input speech under mild noise and reverberation. Track 2 addresses enhancement codecs, which combine coding and compression with denoising and dereverberation. This paper presents the official baseline systems for both tracks in the 2025 LRAC Challenge. The baselines are convolutional neural codec models with Residual Vector Quantization, trained end-to-end using a combination of adversarial and reconstruction objectives. We detail the data filtering and augmentation strategies, model architectures, optimization procedures, and checkpoint selection criteria.

comment: Low-Resource Audio Codec Challenge 2025

♻ ☆ SonicMaster: Towards Controllable All-in-One Music Restoration and Mastering

Jan Melechovsky, Ambuj Mehrish, Abhinaba Roy, Dorien Herremans

Music recordings often suffer from audio quality issues such as excessive reverberation, distortion, clipping, tonal imbalances, and a narrowed stereo image, especially when created in non-professional settings without specialized equipment or expertise. These problems are typically corrected using separate specialized tools and manual adjustments. In this paper, we introduce SonicMaster, the first unified generative model for music restoration and mastering that addresses a broad spectrum of audio artifacts with text-based control. SonicMaster is conditioned on natural language instructions to apply targeted enhancements, or can operate in an automatic mode for general restoration. To train this model, we construct the SonicMaster dataset, a large dataset of paired degraded and high-quality tracks by simulating common degradation types with nineteen degradation functions belonging to five enhancements groups: equalization, dynamics, reverb, amplitude, and stereo. Our approach leverages a flow-matching generative training paradigm to learn an audio transformation that maps degraded inputs to their cleaned, mastered versions guided by text prompts. Objective audio quality metrics demonstrate that SonicMaster significantly improves sound quality across all artifact categories. Furthermore, subjective listening tests confirm that listeners prefer SonicMaster's enhanced outputs over the original degraded audio, highlighting the effectiveness of our unified approach.

☆ MuFFIN: Multifaceted Pronunciation Feedback Model with Interactive Hierarchical Neural Modeling

Bi-Cheng Yan, Ming-Kang Tsai, Berlin Chen

Computer-assisted pronunciation training (CAPT) manages to facilitate second-language (L2) learners to practice pronunciation skills by offering timely and instructive feedback. To examine pronunciation proficiency from multiple facets, existing methods for CAPT broadly fall into two categories: mispronunciation detection and diagnosis (MDD) as well as automatic pronunciation assessment (APA). The former aims to pinpoint phonetic pronunciation errors and provide diagnostic feedback, while the latter seeks instead to quantify pronunciation proficiency pertaining to various aspects. Despite the natural complementarity between MDD and APA, researchers and practitioners, however, often treat them as independent tasks with disparate modeling paradigms. In light of this, we in this paper first introduce MuFFIN, a Multi-Faceted pronunciation Feedback model with an Interactive hierarchical Neural architecture, to jointly address the tasks of MDD and APA. To better capture the nuanced distinctions between phonemes in the feature space, a novel phoneme-contrastive ordinal regularization mechanism is then put forward to optimize the proposed model to generate more phoneme-discriminative features while factoring in the ordinality of the aspect scores. In addition, to address the intricate data imbalance problem in MDD, we design a simple yet effective training objective, which is specifically tailored to perturb the outputs of a phoneme classifier with the phoneme-specific variations, so as to better render the distribution of predicted phonemes meanwhile considering their mispronunciation characteristics. A series of experiments conducted on the Speechocean762 benchmark dataset demonstrates the efficacy of our method in relation to several cutting-edge baselines, showing state-of-the-art performance on both the APA and MDD tasks.

comment: Submitted to IEEE/ACM Transactions on Audio, Speech, and Language Processing

☆ Perceptual Evaluation of Extrapolated Spatial Room Impulse Responses From a Mono Source

Ben Heritage, Fiona Ryder, Michael McLoughlin, Karolina Prawda

Immersion in virtual and augmented reality solutions is reliant on plausible spatial audio. However, plausibly representing a space for immersive audio often requires many individual acoustic measurements of source-microphone pairs with specialist spatial microphones, making the procedure time-consuming and expensive. In this study, we evaluate the plausibility of extrapolated and spatialised Room Impulse Responses (RIRs) by using a 3-Alternative Forced Choice (3AFC) listening test. The stimuli comprised of RIRs from three spaces convolved with speech, orchestral, and instrumental music. When asked to select which stimuli was artificial out of one extrapolated and two real stimuli, an overall accuracy of 38% was achieved from 20 participants (5 percentage points above the expected guessing rate). Given the listening test result, this study shows that it is possible to extrapolate plausible spatial RIRs from mono measurements, decreasing the need for time and specialist equipment in acoustic measurements.

comment: Preprint to be presented as a poster at ADC 2025

☆ AURA Score: A Metric For Holistic Audio Question Answering Evaluation

Satvik Dixit, Soham Deshmukh, Bhiksha Raj

Audio Question Answering (AQA) is a key task for evaluating Audio-Language Models (ALMs), yet assessing open-ended responses remains challenging. Existing metrics used for AQA such as BLEU, METEOR and BERTScore, mostly adapted from NLP and audio captioning, rely on surface similarity and fail to account for question context, reasoning, and partial correctness. To address the gap in literature, we make three contributions in this work. First, we introduce AQEval to enable systematic benchmarking of AQA metrics. It is the first benchmark of its kind, consisting of 10k model responses annotated by multiple humans for their correctness and relevance. Second, we conduct a comprehensive analysis of existing AQA metrics on AQEval, highlighting weak correlation with human judgment, especially for longer answers. Third, we propose a new metric - AURA score, to better evaluate open-ended model responses. On AQEval, AURA achieves state-of-the-art correlation with human ratings, significantly outperforming all baselines. Through this work, we aim to highlight the limitations of current AQA evaluation methods and motivate better metrics. We release both the AQEval benchmark and the AURA metric to support future research in holistic AQA evaluation.

☆ Speak, Edit, Repeat: High-Fidelity Voice Editing and Zero-Shot TTS with Cross-Attentive Mamba

Baher Mohammad, Magauiya Zhussip, Stamatios Lefkimmiatis

We introduce MAVE (Mamba with Cross-Attention for Voice Editing and Synthesis), a novel autoregressive architecture for text-conditioned voice editing and high-fidelity text-to-speech (TTS) synthesis, built on a cross-attentive Mamba backbone. MAVE achieves state-of-the-art performance in speech editing and very competitive results in zero-shot TTS, while not being explicitly trained on the latter task, outperforming leading autoregressive and diffusion models on diverse, real-world audio. By integrating Mamba for efficient audio sequence modeling with cross-attention for precise text-acoustic alignment, MAVE enables context-aware voice editing with exceptional naturalness and speaker consistency. In pairwise human evaluations on a random 40-sample subset of the RealEdit benchmark (400 judgments), 57.2% of listeners rated MAVE - edited speech as perceptually equal to the original, while 24.8% prefered the original and 18.0% MAVE - demonstrating that in the majority of cases edits are indistinguishable from the source. MAVE compares favorably with VoiceCraft and FluentSpeech both on pairwise comparisons and standalone mean opinion score (MOS) evaluations. For zero-shot TTS, MAVE exceeds VoiceCraft in both speaker similarity and naturalness, without requiring multiple inference runs or post-processing. Remarkably, these quality gains come with a significantly lower memory cost and approximately the same latency: MAVE requires ~6x less memory than VoiceCraft during inference on utterances from the RealEdit database (mean duration: 6.21s, A100, FP16, batch size 1). Our results demonstrate that MAVE establishes a new standard for flexible, high-fidelity voice editing and synthesis through the synergistic integration of structured state-space modeling and cross-modal attention.

☆ UniVoice: Unifying Autoregressive ASR and Flow-Matching based TTS with Large Language Models

Wenhao Guan, Zhikang Niu, Ziyue Jiang, Kaidi Wang, Peijie Chen, Qingyang Hong, Lin Li, Xie Chen

Large language models (LLMs) have demonstrated promising performance in both automatic speech recognition (ASR) and text-to-speech (TTS) systems, gradually becoming the mainstream approach. However, most current approaches address these tasks separately rather than through a unified framework. This work aims to integrate these two tasks into one unified model. Although discrete speech tokenization enables joint modeling, its inherent information loss limits performance in both recognition and generation. In this work, we present UniVoice, a unified LLM framework through continuous representations that seamlessly integrates speech recognition and synthesis within a single model. Our approach combines the strengths of autoregressive modeling for speech recognition with flow matching for high-quality generation. To mitigate the inherent divergence between autoregressive and flow-matching models, we further design a dual attention mechanism, which switches between a causal mask for recognition and a bidirectional attention mask for synthesis. Furthermore, the proposed text-prefix-conditioned speech infilling method enables high-fidelity zero-shot voice cloning. Experimental results demonstrate that our method can achieve or exceed current single-task modeling methods in both ASR and zero-shot TTS tasks. This work explores new possibilities for end-to-end speech understanding and generation.

☆ Robustness assessment of large audio language models in multiple-choice evaluation ICASSP 2026

Fernando López, Santosh Kesiraju, Jordi Luque

Recent advances in large audio language models (LALMs) have primarily been assessed using a multiple-choice question answering (MCQA) framework. However, subtle changes, such as shifting the order of choices, result in substantially different results. Existing MCQA frameworks do not account for this variability and report a single accuracy number per benchmark or category. We dive into the MCQA evaluation framework and conduct a systematic study spanning three benchmarks (MMAU, MMAR and MMSU) and four models: Audio Flamingo 2, Audio Flamingo 3, Qwen2.5-Omni-7B-Instruct, and Kimi-Audio-7B-Instruct. Our findings indicate that models are sensitive not only to the ordering of choices, but also to the paraphrasing of the question and the choices. Finally, we propose a simpler evaluation protocol and metric that account for subtle variations and provide a more detailed evaluation report of LALMs within the MCQA framework.

comment: Submitted to ICASSP 2026

☆ Language Model Based Text-to-Audio Generation: Anti-Causally Aligned Collaborative Residual Transformers EMNLP 2025

Juncheng Wang, Chao Xu, Cheng Yu, Zhe Hu, Haoyu Xie, Guoqi Yu, Lei Shang, Shujun Wang

While language models (LMs) paired with residual vector quantization (RVQ) tokenizers have shown promise in text-to-audio (T2A) generation, they still lag behind diffusion-based models by a non-trivial margin. We identify a critical dilemma underpinning this gap: incorporating more RVQ layers improves audio reconstruction fidelity but exceeds the generation capacity of conventional LMs. To address this, we first analyze RVQ dynamics and uncover two key limitations: 1) orthogonality of features across RVQ layers hinders effective LMs training, and 2) descending semantic richness in tokens from deeper RVQ layers exacerbates exposure bias during autoregressive decoding. Based on these insights, we propose Siren, a novel LM-based framework that employs multiple isolated transformers with causal conditioning and anti-causal alignment via reinforcement learning. Extensive experiments demonstrate that Siren outperforms both existing LM-based and diffusion-based T2A systems, achieving state-of-the-art results. By bridging the representational strengths of LMs with the fidelity demands of audio synthesis, our approach repositions LMs as competitive contenders against diffusion models in T2A tasks. Moreover, by aligning audio representations with linguistic structures, Siren facilitates a promising pathway toward unified multi-modal generation frameworks.

comment: Accepted to EMNLP 2025

☆ Evaluating Self-Supervised Speech Models via Text-Based LLMS ASRU 2025

Takashi Maekaku, Keita Goto, Jinchuan Tian, Yusuke Shinohara, Shinji Watanabe

Self-Supervised Learning (SSL) has gained traction for its ability to learn rich representations with low labeling costs, applicable across diverse downstream tasks. However, assessing the downstream-task performance remains challenging due to the cost of extra training and evaluation. Existing methods for task-agnostic evaluation also require extra training or hyperparameter tuning. We propose a novel evaluation metric using large language models (LLMs). By inputting discrete token sequences and minimal domain cues derived from SSL models into LLMs, we obtain the mean log-likelihood; these cues guide in-context learning, rendering the score more reliable without extra training or hyperparameter tuning. Experimental results show a correlation between LLM-based scores and automatic speech recognition task. Additionally, our findings reveal that LLMs not only functions as an SSL evaluation tools but also provides inference-time embeddings that are useful for speaker verification task.

comment: Accepted to ASRU 2025

☆ Differentiable physics for sound field reconstruction

Samuel A. Verburg, Efren Fernandez-Grande, Peter Gerstoft

Sound field reconstruction involves estimating sound fields from a limited number of spatially distributed observations. This work introduces a differentiable physics approach for sound field reconstruction, where the initial conditions of the wave equation are approximated with a neural network, and the differential operator is computed with a differentiable numerical solver. The use of a numerical solver enables a stable network training while enforcing the physics as a strong constraint, in contrast to conventional physics-informed neural networks, which include the physics as a constraint in the loss function. We introduce an additional sparsity-promoting constraint to achieve meaningful solutions even under severe undersampling conditions. Experiments demonstrate that the proposed approach can reconstruct sound fields under extreme data scarcity, achieving higher accuracy and better convergence compared to physics-informed neural networks.

comment: 28 pages plus references, 8 figures, full journal paper

☆ WaveSP-Net: Learnable Wavelet-Domain Sparse Prompt Tuning for Speech Deepfake Detection ICASSP 2026

Xi Xuan, Xuechen Liu, Wenxin Zhang, Yi-Cheng Lin, Xiaojian Lin, Tomi Kinnunen

Modern front-end design for speech deepfake detection relies on full fine-tuning of large pre-trained models like XLSR. However, this approach is not parameter-efficient and may lead to suboptimal generalization to realistic, in-the-wild data types. To address these limitations, we introduce a new family of parameter-efficient front-ends that fuse prompt-tuning with classical signal processing transforms. These include FourierPT-XLSR, which uses the Fourier Transform, and two variants based on the Wavelet Transform: WSPT-XLSR and Partial-WSPT-XLSR. We further propose WaveSP-Net, a novel architecture combining a Partial-WSPT-XLSR front-end and a bidirectional Mamba-based back-end. This design injects multi-resolution features into the prompt embeddings, which enhances the localization of subtle synthetic artifacts without altering the frozen XLSR parameters. Experimental results demonstrate that WaveSP-Net outperforms several state-of-the-art models on two new and challenging benchmarks, Deepfake-Eval-2024 and SpoofCeleb, with low trainable parameters and notable performance gains. The code and models are available at https://github.com/xxuan-acoustics/WaveSP-Net.

comment: Submitted to ICASSP 2026

☆ BaldWhisper: Faster Whisper with Head Shearing and Layer Merging

Yaya Sy, Christophe Cerisara, Irina Illina

Pruning large pre-trained transformers for low-resource languages is challenging, as it often requires massive retraining data to recover performance. For instance, Distill-Whisper prunes Whisper by 40% and retrains on 21,000 hours of speech, far beyond what is available for most languages. Can Whisper be made lighter and faster for edge devices in data-scarce settings? Focusing on Bambara with only 32h of speech-to-text data, we propose a new pruning recipe. Instead of vocabulary pruning, which is unsuitable due to frequent code-switching by Bambara speakers, we compress the embeddings with low-rank decomposition and feature distillation. Rather than removing layers, we merge them to limit performance loss. The final model preserves 90% of the original performance while being 48% smaller and 2.15x faster on a MacBook Air M1.

♻ ☆ TalkPlay-Tools: Conversational Music Recommendation with LLM Tool Calling NeurIPS

Seungheon Doh, Keunwoo Choi, Juhan Nam

While the recent developments in large language models (LLMs) have successfully enabled generative recommenders with natural language interactions, their recommendation behavior is limited, leaving other simpler yet crucial components such as metadata or attribute filtering underutilized in the system. We propose an LLM-based music recommendation system with tool calling to serve as a unified retrieval-reranking pipeline. Our system positions an LLM as an end-to-end recommendation system that interprets user intent, plans tool invocations, and orchestrates specialized components: boolean filters (SQL), sparse retrieval (BM25), dense retrieval (embedding similarity), and generative retrieval (semantic IDs). Through tool planning, the system predicts which types of tools to use, their execution order, and the arguments needed to find music matching user preferences, supporting diverse modalities while seamlessly integrating multiple database filtering methods. We demonstrate that this unified tool-calling framework achieves competitive performance across diverse recommendation scenarios by selectively employing appropriate retrieval methods based on user queries, envisioning a new paradigm for conversational music recommendation systems.

comment: Accepted for publication at The Workshop on AI for Music, Neural Information Processing Systems (NeurIPS-AI4Music)

♻ ☆ EmoSSLSphere: Multilingual Emotional Speech Synthesis with Spherical Vectors and Discrete Speech Tokens ISCA

Joonyong Park, Kenichi Nakamura

This paper introduces EmoSSLSphere, a novel framework for multilingual emotional text-to-speech (TTS) synthesis that combines spherical emotion vectors with discrete token features derived from self-supervised learning (SSL). By encoding emotions in a continuous spherical coordinate space and leveraging SSL-based representations for semantic and acoustic modeling, EmoSSLSphere enables fine-grained emotional control, effective cross-lingual emotion transfer, and robust preservation of speaker identity. We evaluate EmoSSLSphere on English and Japanese corpora, demonstrating significant improvements in speech intelligibility, spectral fidelity, prosodic consistency, and overall synthesis quality. Subjective evaluations further confirm that our method outperforms baseline models in terms of naturalness and emotional expressiveness, underscoring its potential as a scalable solution for multilingual emotional TTS.

comment: In Proceedings of the 13th ISCA Speech Synthesis Workshop

♻ ☆ SonicMaster: Towards Controllable All-in-One Music Restoration and Mastering

Jan Melechovsky, Ambuj Mehrish, Abhinaba Roy, Dorien Herremans

Music recordings often suffer from audio quality issues such as excessive reverberation, distortion, clipping, tonal imbalances, and a narrowed stereo image, especially when created in non-professional settings without specialized equipment or expertise. These problems are typically corrected using separate specialized tools and manual adjustments. In this paper, we introduce SonicMaster, the first unified generative model for music restoration and mastering that addresses a broad spectrum of audio artifacts with text-based control. SonicMaster is conditioned on natural language instructions to apply targeted enhancements, or can operate in an automatic mode for general restoration. To train this model, we construct the SonicMaster dataset, a large dataset of paired degraded and high-quality tracks by simulating common degradation types with nineteen degradation functions belonging to five enhancements groups: equalization, dynamics, reverb, amplitude, and stereo. Our approach leverages a flow-matching generative training paradigm to learn an audio transformation that maps degraded inputs to their cleaned, mastered versions guided by text prompts. Objective audio quality metrics demonstrate that SonicMaster significantly improves sound quality across all artifact categories. Furthermore, subjective listening tests confirm that listeners prefer SonicMaster's enhanced outputs over the original degraded audio, highlighting the effectiveness of our unified approach.

♻ ☆ Multilingual Dataset Integration Strategies for Robust Audio Deepfake Detection: A SAFE Challenge System ASRU 2025

Hashim Ali, Surya Subramani, Lekha Bollinani, Nithin Sai Adupa, Sali El-Loh, Hafiz Malik

The SAFE Challenge evaluates synthetic speech detection across three tasks: unmodified audio, processed audio with compression artifacts, and laundered audio designed to evade detection. We systematically explore self-supervised learning (SSL) front-ends, training data compositions, and audio length configurations for robust deepfake detection. Our AASIST-based approach incorporates WavLM large frontend with RawBoost augmentation, trained on a multilingual dataset of 256,600 samples spanning 9 languages and over 70 TTS systems from CodecFake, MLAAD v5, SpoofCeleb, Famous Figures, and MAILABS. Through extensive experimentation with different SSL front-ends, three training data versions, and two audio lengths, we achieved second place in both Task 1 (unmodified audio detection) and Task 3 (laundered audio detection), demonstrating strong generalization and robustness.

comment: Accepted @ IEEE ASRU 2025