Eng Siong Chng

Chen Chen, Yuchen Hu, Chao-Han Huck Yang et al. · gatech

Advancements in deep neural networks have allowed automatic speech recognition (ASR) systems to attain human parity on several publicly available clean speech datasets. However, even state-of-the-art ASR systems experience performance degradation when confronted with adverse conditions, as a well-trained acoustic model is sensitive to variations in the speech domain, e.g., background noise. Intuitively, humans address this issue by relying on their linguistic knowledge: the meaning of ambiguous spoken terms is usually inferred from contextual cues thereby reducing the dependency on the auditory system. Inspired by this observation, we introduce the first open-source benchmark to utilize external large language models (LLMs) for ASR error correction, where N-best decoding hypotheses provide informative elements for true transcription prediction. This approach is a paradigm shift from the traditional language model rescoring strategy that can only select one candidate hypothesis as the output transcription. The proposed benchmark contains a novel dataset, HyPoradise (HP), encompassing more than 334,000 pairs of N-best hypotheses and corresponding accurate transcriptions across prevalent speech domains. Given this dataset, we examine three types of error correction techniques based on LLMs with varying amounts of labeled hypotheses-transcription pairs, which gains a significant word error rate (WER) reduction. Experimental evidence demonstrates the proposed technique achieves a breakthrough by surpassing the upper bound of traditional re-ranking based methods. More surprisingly, LLM with reasonable prompt and its generative capability can even correct those tokens that are missing in N-best list. We make our results publicly accessible for reproducible pipelines with released pre-trained models, thus providing a new evaluation paradigm for ASR error correction with LLMs.

Heqing Zou, Yuke Si, Chen Chen et al.

Speech Emotion Recognition (SER) aims to help the machine to understand human's subjective emotion from only audio information. However, extracting and utilizing comprehensive in-depth audio information is still a challenging task. In this paper, we propose an end-to-end speech emotion recognition system using multi-level acoustic information with a newly designed co-attention module. We firstly extract multi-level acoustic information, including MFCC, spectrogram, and the embedded high-level acoustic information with CNN, BiLSTM and wav2vec2, respectively. Then these extracted features are treated as multimodal inputs and fused by the proposed co-attention mechanism. Experiments are carried on the IEMOCAP dataset, and our model achieves competitive performance with two different speaker-independent cross-validation strategies. Our code is available on GitHub.

Yuchen Hu, Chen Chen, Heqing Zou et al.

Recent studies in neural network-based monaural speech separation (SS) have achieved a remarkable success thanks to increasing ability of long sequence modeling. However, they would degrade significantly when put under realistic noisy conditions, as the background noise could be mistaken for speaker's speech and thus interfere with the separated sources. To alleviate this problem, we propose a novel network to unify speech enhancement and separation with gradient modulation to improve noise-robustness. Specifically, we first build a unified network by combining speech enhancement (SE) and separation modules, with multi-task learning for optimization, where SE is supervised by parallel clean mixture to reduce noise for downstream speech separation. Furthermore, in order to avoid suppressing valid speaker information when reducing noise, we propose a gradient modulation (GM) strategy to harmonize the SE and SS tasks from optimization view. Experimental results show that our approach achieves the state-of-the-art on large-scale Libri2Mix- and Libri3Mix-noisy datasets, with SI-SNRi results of 16.0 dB and 15.8 dB respectively. Our code is available at GitHub.

Chao-Han Huck Yang, Taejin Park, Yuan Gong et al. · gatech

Given recent advances in generative AI technology, a key question is how large language models (LLMs) can enhance acoustic modeling tasks using text decoding results from a frozen, pretrained automatic speech recognition (ASR) model. To explore new capabilities in language modeling for speech processing, we introduce the generative speech transcription error correction (GenSEC) challenge. This challenge comprises three post-ASR language modeling tasks: (i) post-ASR transcription correction, (ii) speaker tagging, and (iii) emotion recognition. These tasks aim to emulate future LLM-based agents handling voice-based interfaces while remaining accessible to a broad audience by utilizing open pretrained language models or agent-based APIs. We also discuss insights from baseline evaluations, as well as lessons learned for designing future evaluations.

Yuchen Hu, Chen Chen, Ruizhe Li et al.

Speech enhancement (SE) is proved effective in reducing noise from noisy speech signals for downstream automatic speech recognition (ASR), where multi-task learning strategy is employed to jointly optimize these two tasks. However, the enhanced speech learned by SE objective may not always yield good ASR results. From the optimization view, there sometimes exists interference between the gradients of SE and ASR tasks, which could hinder the multi-task learning and finally lead to sub-optimal ASR performance. In this paper, we propose a simple yet effective approach called gradient remedy (GR) to solve interference between task gradients in noise-robust speech recognition, from perspectives of both angle and magnitude. Specifically, we first project the SE task's gradient onto a dynamic surface that is at acute angle to ASR gradient, in order to remove the conflict between them and assist in ASR optimization. Furthermore, we adaptively rescale the magnitude of two gradients to prevent the dominant ASR task from being misled by SE gradient. Experimental results show that the proposed approach well resolves the gradient interference and achieves relative word error rate (WER) reductions of 9.3% and 11.1% over multi-task learning baseline, on RATS and CHiME-4 datasets, respectively. Our code is available at GitHub.

Shangeth Rajaa, Kriti Anandan, Swaraj Dalmia et al.

The pre-trained multi-lingual XLSR model generalizes well for language identification after fine-tuning on unseen languages. However, the performance significantly degrades when the languages are not very distinct from each other, for example, in the case of dialects. Low resource dialect classification remains a challenging problem to solve. We present a new data augmentation method that leverages model training dynamics of individual data points to improve sampling for latent mixup. The method works well in low-resource settings where generalization is paramount. Our datamaps-based mixup technique, which we call Map-Mix improves weighted F1 scores by 2% compared to the random mixup baseline and results in a significantly well-calibrated model. The code for our method is open sourced on https://github.com/skit-ai/Map-Mix.

Sathya Krishnan Suresh, Wu Mengjun, Tushar Pranav et al.

The scarcity of domain-specific dialogue datasets limits the development of dialogue systems across applications. Existing research is constrained by general or niche datasets that lack sufficient scale for training dialogue systems. To address this gap, we introduce DiaSynth - a synthetic dialogue generation framework capable of generating high-quality, contextually rich dialogues across a wide range of domains. Unlike existing frameworks, DiaSynth uses Large Language Models (LLMs) and Chain of Thought (CoT) reasoning to generate dynamic, domain-specific dialogues with simulated personas and diverse conversational features. We perform our experiments by generating synthetic data using different LLMs and few-shot examples from DialogSum and SAMSum. The pretrained language models fine-tuned on the synthetic data outperform the base models by 16.47% on dialogue summarization, while the comparison between models fine-tuned on in-domain data and synthetic data shows that the synthetic data is able to capture 90.48% of the performance distribution of the in-domain data on dialogue summarization. The quality of the data generated also increases as we increase the size of LLM from 3B to 8B. These results validate DiaSynth's potential as a robust alternative to traditional data collection methods. We open source the code and data generated for future research.

Yuchen Hu, Chen Chen, Ruizhe Li et al.

With recent advances of diffusion model, generative speech enhancement (SE) has attracted a surge of research interest due to its great potential for unseen testing noises. However, existing efforts mainly focus on inherent properties of clean speech, underexploiting the varying noise information in real world. In this paper, we propose a noise-aware speech enhancement (NASE) approach that extracts noise-specific information to guide the reverse process in diffusion model. Specifically, we design a noise classification (NC) model to produce acoustic embedding as a noise conditioner to guide the reverse denoising process. Meanwhile, a multi-task learning scheme is devised to jointly optimize SE and NC tasks to enhance the noise specificity of conditioner. NASE is shown to be a plug-and-play module that can be generalized to any diffusion SE models. Experiments on VB-DEMAND dataset show that NASE effectively improves multiple mainstream diffusion SE models, especially on unseen noises.

Yuchen Hu, Chen Chen, Qiushi Zhu et al.

Automatic speech recognition (ASR) has gained remarkable successes thanks to recent advances of deep learning, but it usually degrades significantly under real-world noisy conditions. Recent works introduce speech enhancement (SE) as front-end to improve speech quality, which is proved effective but may not be optimal for downstream ASR due to speech distortion problem. Based on that, latest works combine SE and currently popular self-supervised learning (SSL) to alleviate distortion and improve noise robustness. Despite the effectiveness, the speech distortion caused by conventional SE still cannot be cleared out. In this paper, we propose a self-supervised framework named Wav2code to implement a feature-level SE with reduced distortions for noise-robust ASR. First, in pre-training stage the clean speech representations from SSL model are sent to lookup a discrete codebook via nearest-neighbor feature matching, the resulted code sequence are then exploited to reconstruct the original clean representations, in order to store them in codebook as prior. Second, during finetuning we propose a Transformer-based code predictor to accurately predict clean codes by modeling the global dependency of input noisy representations, which enables discovery and restoration of high-quality clean representations with reduced distortions. Furthermore, we propose an interactive feature fusion network to combine original noisy and the restored clean representations to consider both fidelity and quality, resulting in more informative features for downstream ASR. Finally, experiments on both synthetic and real noisy datasets demonstrate that Wav2code can solve the speech distortion and improve ASR performance under various noisy conditions, resulting in stronger robustness.

Yuchen Hu, Ruizhe Li, Chen Chen et al.

Audio-visual speech recognition (AVSR) provides a promising solution to ameliorate the noise-robustness of audio-only speech recognition with visual information. However, most existing efforts still focus on audio modality to improve robustness considering its dominance in AVSR task, with noise adaptation techniques such as front-end denoise processing. Though effective, these methods are usually faced with two practical challenges: 1) lack of sufficient labeled noisy audio-visual training data in some real-world scenarios and 2) less optimal model generality to unseen testing noises. In this work, we investigate the noise-invariant visual modality to strengthen robustness of AVSR, which can adapt to any testing noises while without dependence on noisy training data, a.k.a., unsupervised noise adaptation. Inspired by human perception mechanism, we propose a universal viseme-phoneme mapping (UniVPM) approach to implement modality transfer, which can restore clean audio from visual signals to enable speech recognition under any noisy conditions. Extensive experiments on public benchmarks LRS3 and LRS2 show that our approach achieves the state-of-the-art under various noisy as well as clean conditions. In addition, we also outperform previous state-of-the-arts on visual speech recognition task.

Chen Chen, Yuchen Hu, Chao-Han Huck Yang et al. · gatech

Code-switching (CS) speech refers to the phenomenon of mixing two or more languages within the same sentence. Despite the recent advances in automatic speech recognition (ASR), CS-ASR is still a challenging task ought to the grammatical structure complexity of the phenomenon and the data scarcity of specific training corpus. In this work, we propose to leverage large language models (LLMs) and lists of hypotheses generated by an ASR to address the CS problem. Specifically, we first employ multiple well-trained ASR models for N-best hypotheses generation, with the aim of increasing the diverse and informative elements in the set of hypotheses. Next, we utilize the LLMs to learn the hypotheses-to-transcription (H2T) mapping by adding a trainable low-rank adapter. Such a generative error correction (GER) method directly predicts the accurate transcription according to its expert linguistic knowledge and N-best hypotheses, resulting in a paradigm shift from the traditional language model rescoring or error correction techniques. Experimental evidence demonstrates that GER significantly enhances CS-ASR accuracy, in terms of reduced mixed error rate (MER). Furthermore, LLMs show remarkable data efficiency for H2T learning, providing a potential solution to the data scarcity problem of CS-ASR in low-resource languages.

Chen Chen, Nana Hou, Yuchen Hu et al.

Noise-robust speech recognition systems require large amounts of training data including noisy speech data and corresponding transcripts to achieve state-of-the-art performances in face of various practical environments. However, such plenty of in-domain data is not always available in the real-life world. In this paper, we propose a generative adversarial network to simulate noisy spectrum from the clean spectrum (Simu-GAN), where only 10 minutes of unparalleled in-domain noisy speech data is required as labels. Furthermore, we also propose a dual-path speech recognition system to improve the robustness of the system under noisy conditions. Experimental results show that the proposed speech recognition system achieves 7.3% absolute improvement with simulated noisy data by Simu-GAN over the best baseline in terms of word error rate (WER).

Yang Xiao, Xubo Liu, James King et al.

Continuously learning new classes without catastrophic forgetting is a challenging problem for on-device environmental sound classification given the restrictions on computation resources (e.g., model size, running memory). To address this issue, we propose a simple and efficient continual learning method. Our method selects the historical data for the training by measuring the per-sample classification uncertainty. Specifically, we measure the uncertainty by observing how the classification probability of data fluctuates against the parallel perturbations added to the classifier embedding. In this way, the computation cost can be significantly reduced compared with adding perturbation to the raw data. Experimental results on the DCASE 2019 Task 1 and ESC-50 dataset show that our proposed method outperforms baseline continual learning methods on classification accuracy and computational efficiency, indicating our method can efficiently and incrementally learn new classes without the catastrophic forgetting problem for on-device environmental sound classification.

Chen Chen, Nana Hou, Yuchen Hu et al.

Automated Audio captioning (AAC) is a cross-modal task that generates natural language to describe the content of input audio. Most prior works usually extract single-modality acoustic features and are therefore sub-optimal for the cross-modal decoding task. In this work, we propose a novel AAC system called CLIP-AAC to learn interactive cross-modality representation with both acoustic and textual information. Specifically, the proposed CLIP-AAC introduces an audio-head and a text-head in the pre-trained encoder to extract audio-text information. Furthermore, we also apply contrastive learning to narrow the domain difference by learning the correspondence between the audio signal and its paired captions. Experimental results show that the proposed CLIP-AAC approach surpasses the best baseline by a significant margin on the Clotho dataset in terms of NLP evaluation metrics. The ablation study indicates that both the pre-trained model and contrastive learning contribute to the performance gain of the AAC model.

Chen Chen, Yuchen Hu, Weiwei Weng et al.

Deep neural network based speech enhancement technique focuses on learning a noisy-to-clean transformation supervised by paired training data. However, the task-specific evaluation metric (e.g., PESQ) is usually non-differentiable and can not be directly constructed in the training criteria. This mismatch between the training objective and evaluation metric likely results in sub-optimal performance. To alleviate it, we propose a metric-oriented speech enhancement method (MOSE), which leverages the recent advances in the diffusion probabilistic model and integrates a metric-oriented training strategy into its reverse process. Specifically, we design an actor-critic based framework that considers the evaluation metric as a posterior reward, thus guiding the reverse process to the metric-increasing direction. The experimental results demonstrate that MOSE obviously benefits from metric-oriented training and surpasses the generative baselines in terms of all evaluation metrics.

Yuchen Hu, Nana Hou, Chen Chen et al.

Automatic speech recognition (ASR) systems degrade significantly under noisy conditions. Recently, speech enhancement (SE) is introduced as front-end to reduce noise for ASR, but it also suppresses some important speech information, i.e., over-suppression. To alleviate this, we propose a dual-path style learning approach for end-to-end noise-robust speech recognition (DPSL-ASR). Specifically, we first introduce clean speech feature along with the fused feature from IFF-Net as dual-path inputs to recover the suppressed information. Then, we propose style learning to map the fused feature close to clean feature, in order to learn latent speech information from the latter, i.e., clean "speech style". Furthermore, we also minimize the distance of final ASR outputs in two paths to improve noise-robustness. Experiments show that the proposed approach achieves relative word error rate (WER) reductions of 10.6% and 8.6% over the best IFF-Net baseline, on RATS and CHiME-4 datasets respectively.

Chen Chen, Yuchen Hu, Nana Hou et al.

Although automatic speech recognition (ASR) task has gained remarkable success by sequence-to-sequence models, there are two main mismatches between its training and testing that might lead to performance degradation: 1) The typically used cross-entropy criterion aims to maximize log-likelihood of the training data, while the performance is evaluated by word error rate (WER), not log-likelihood; 2) The teacher-forcing method leads to the dependence on ground truth during training, which means that model has never been exposed to its own prediction before testing. In this paper, we propose an optimization method called self-critical sequence training (SCST) to make the training procedure much closer to the testing phase. As a reinforcement learning (RL) based method, SCST utilizes a customized reward function to associate the training criterion and WER. Furthermore, it removes the reliance on teacher-forcing and harmonizes the model with respect to its inference procedure. We conducted experiments on both clean and noisy speech datasets, and the results show that the proposed SCST respectively achieves 8.7% and 7.8% relative improvements over the baseline in terms of WER.

Chen Chen, Yuchen Hu, Heqing Zou et al.

Deep neural network based speech enhancement approaches aim to learn a noisy-to-clean transformation using a supervised learning paradigm. However, such a trained-well transformation is vulnerable to unseen noises that are not included in training set. In this work, we focus on the unsupervised noise adaptation problem in speech enhancement, where the ground truth of target domain data is completely unavailable. Specifically, we propose a generative adversarial network based method to efficiently learn a converse clean-to-noisy transformation using a few minutes of unpaired target domain data. Then this transformation is utilized to generate sufficient simulated data for domain adaptation of the enhancement model. Experimental results show that our method effectively mitigates the domain mismatch between training and test sets, and surpasses the best baseline by a large margin.

Yuchen Hu, Chen Chen, Ruizhe Li et al.

Audio-visual speech recognition (AVSR) attracts a surge of research interest recently by leveraging multimodal signals to understand human speech. Mainstream approaches addressing this task have developed sophisticated architectures and techniques for multi-modality fusion and representation learning. However, the natural heterogeneity of different modalities causes distribution gap between their representations, making it challenging to fuse them. In this paper, we aim to learn the shared representations across modalities to bridge their gap. Different from existing similar methods on other multimodal tasks like sentiment analysis, we focus on the temporal contextual dependencies considering the sequence-to-sequence task setting of AVSR. In particular, we propose an adversarial network to refine frame-level modality-invariant representations (MIR-GAN), which captures the commonality across modalities to ease the subsequent multimodal fusion process. Extensive experiments on public benchmarks LRS3 and LRS2 show that our approach outperforms the state-of-the-arts.

Yizhou Peng, Yufei Liu, Jicheng Zhang et al.

Internal Language Model Estimation (ILME) based language model (LM) fusion has been shown significantly improved recognition results over conventional shallow fusion in both intra-domain and cross-domain speech recognition tasks. In this paper, we attempt to apply our ILME method to cross-domain code-switching speech recognition (CSSR) work. Specifically, our curiosity comes from several aspects. First, we are curious about how effective the ILME-based LM fusion is for both intra-domain and cross-domain CSSR tasks. We verify this with or without merging two code-switching domains. More importantly, we train an end-to-end (E2E) speech recognition model by means of merging two monolingual data sets and observe the efficacy of the proposed ILME-based LM fusion for CSSR. Experimental results on SEAME that is from Southeast Asian and another Chinese Mainland CS data set demonstrate the effectiveness of the proposed ILME-based LM fusion method.

Chen Chen, Yuchen Hu, Qiang Zhang et al.

Audio-visual speech recognition (AVSR) has gained remarkable success for ameliorating the noise-robustness of speech recognition. Mainstream methods focus on fusing audio and visual inputs to obtain modality-invariant representations. However, such representations are prone to over-reliance on audio modality as it is much easier to recognize than video modality in clean conditions. As a result, the AVSR model underestimates the importance of visual stream in face of noise corruption. To this end, we leverage visual modality-specific representations to provide stable complementary information for the AVSR task. Specifically, we propose a reinforcement learning (RL) based framework called MSRL, where the agent dynamically harmonizes modality-invariant and modality-specific representations in the auto-regressive decoding process. We customize a reward function directly related to task-specific metrics (i.e., word error rate), which encourages the MSRL to effectively explore the optimal integration strategy. Experimental results on the LRS3 dataset show that the proposed method achieves state-of-the-art in both clean and various noisy conditions. Furthermore, we demonstrate the better generality of MSRL system than other baselines when test set contains unseen noises.

Alexey Sholokhov, Nikita Kuzmin, Kong Aik Lee et al.

This paper focuses on multi-enrollment speaker recognition which naturally occurs in the task of online speaker clustering, and studies the properties of different scoring back-ends in this scenario. First, we show that popular cosine scoring suffers from poor score calibration with a varying number of enrollment utterances. Second, we propose a simple replacement for cosine scoring based on an extremely constrained version of probabilistic linear discriminant analysis (PLDA). The proposed model improves over the cosine scoring for multi-enrollment recognition while keeping the same performance in the case of one-to-one comparisons. Finally, we consider an online speaker clustering task where each step naturally involves multi-enrollment recognition. We propose an online clustering algorithm allowing us to take benefits from the PLDA model such as the ability to handle uncertainty and better score calibration. Our experiments demonstrate the effectiveness of the proposed algorithm.

Andrew Koh, Eng Siong Chng

In this paper, we tackle the new Language-Based Audio Retrieval task proposed in DCASE 2022. Firstly, we introduce a simple, scalable architecture which ties both the audio and text encoder together. Secondly, we show that using this architecture along with contrastive loss allows the model to significantly beat the performance of the baseline model. Finally, in addition to having an extremely low training memory requirement, we are able to use pretrained models as it is without needing to finetune them. We test our methods and show that using a combination of our methods beats the baseline scores significantly.

Yuhang Yang, Haihua Xu, Hao Huang et al.

To let the state-of-the-art end-to-end ASR model enjoy data efficiency, as well as much more unpaired text data by multi-modal training, one needs to address two problems: 1) the synchronicity of feature sampling rates between speech and language (aka text data); 2) the homogeneity of the learned representations from two encoders. In this paper we propose to employ a novel bidirectional attention mechanism (BiAM) to jointly learn both ASR encoder (bottom layers) and text encoder with a multi-modal learning method. The BiAM is to facilitate feature sampling rate exchange, realizing the quality of the transformed features for the one kind to be measured in another space, with diversified objective functions. As a result, the speech representations are enriched with more linguistic information, while the representations generated by the text encoder are more similar to corresponding speech ones, and therefore the shared ASR models are more amenable for unpaired text data pretraining. To validate the efficacy of the proposed method, we perform two categories of experiments with or without extra unpaired text data. Experimental results on Librispeech corpus show it can achieve up to 6.15% word error rate reduction (WERR) with only paired data learning, while 9.23% WERR when more unpaired text data is employed.

Yuchen Hu, Chen Chen, Siyin Wang et al.

In this paper, we propose reverse inference optimization (RIO), a simple and effective method designed to enhance the robustness of autoregressive-model-based zero-shot text-to-speech (TTS) systems using reinforcement learning from human feedback (RLHF). To assess the quality of speech produced by the TTS system without human annotations, RIO introduces a novel concept termed as reverse inference based on the Bayesian principle, which suggests that a high-quality generated speech should be able to be used as a prompt for subsequent generation using the same TTS model. By leveraging reverse inference as the standard to select exemplars used in RLHF from the speech samples generated by the TTS system itself, RIO steers the subsequent optimization towards a direction of enhancing the TTS robustness. The RIO framework, comprising sampling, automatic annotating, and learning, obviates the need for a reward model or pairwise preference data, and significantly improves the stability of zero-shot TTS performance by reducing the discrepancies between training and inference conditions. Our experimental results verify that RIO can effectively improve both subjective and objective metrics, including mean opinion scores, word error rates, and speaker similarity. Remarkably, RIO can also diminish the incidence of bad outputs to nearly zero percent, rivalling the robustness when using ground-truth speech as the prompt.

Chin Yuen Kwok, Jia Qi Yip, Eng Siong Chng

Continual Learning (CL) involves fine-tuning pre-trained models with new data while maintaining the performance on the pre-trained data. This is particularly relevant for expanding multilingual ASR (MASR) capabilities. However, existing CL methods, mainly designed for computer vision and reinforcement learning tasks, often yield sub-optimal results when directly applied to MASR. We hypothesise that this is because CL of the auto-regressive decoder in the MASR model is difficult. To verify this, we propose four optimizations on the decoder. They include decoder-layer gradient surgery, freezing unused token embeddings, suppressing output of newly added tokens, and learning rate re-scaling. Our experiments on adapting Whisper to 10 unseen languages from the Common Voice dataset demonstrate that these optimizations reduce the Average Word Error Rate (AWER) of pretrained languages from 14.2% to 12.4% compared with Experience Replay, without compromising the AWER of new languages.

Yayue Deng, Guoqiang Hu, Haiyang Sun et al.

Spoken Dialogue Models (SDMs) have advanced rapidly, yet their ability to sustain genuinely interactive multi-turn conversations remains underexplored, as most benchmarks focus on single-turn exchanges. We introduce Multi-Bench, the first benchmark explicitly designed to evaluate SDMs in multi-turn interactive dialogue with an emphasis on emotional intelligence. Multi-Bench employs a hierarchical structure with a basic track for emotion understanding and reasoning and an advanced track for emotion support and application. It comprises five carefully designed tasks and about 3.2K samples, ranging from emotion recognition to complex reasoning and interactive dialogue, supported by a reproducible evaluation framework. We evaluate six representative SDMs on eight subsets of Multi-Bench. Results show that while current SDMs achieve good performance on basic understanding tasks, they still have room for improvement in advanced multi-turn interactive dialogue and reasoning-related tasks, particularly in emotion awareness and application.

Yang Yuhang, Peng Yizhou, Eng Siong Chng et al.

The integration of large language models (LLMs) with pre-trained speech models has opened up new avenues in automatic speech recognition (ASR). While LLMs excel in multimodal understanding tasks, effectively leveraging their capabilities for ASR remains a significant challenge. This paper presents a novel training approach to enhance LLM performance in ASR tasks. We propose pre-training LLMs on Pinyin embedding sequences, which represent pronunciation features, to generate corresponding Chinese characters. This step enables the LLM to adapt to generating text from pronunciation features before encountering real speech data. Furthermore, we fine-tune the LoRA parameters to enhance the LLM's understanding of speech modality information. In AISHELL-1 corpus, our approach yields a 9.5% relative improvement in ASR tasks compared to the baseline without Pinyi-to-Character pre-training. Additionally, incorporating auxiliary text data for Pinyi-to-Character pre-training further boosts performance, achieving a 19.0% relative improvement.

Ashutosh Anshul, Shreyas Gopal, Deepu Rajan et al.

Recent multimodal deepfake detection methods designed for generalization conjecture that single-stage supervised training struggles to generalize across unseen manipulations and datasets. However, such approaches that target generalization require pretraining over real samples. Additionally, these methods primarily focus on detecting audio-visual inconsistencies and may overlook intra-modal artifacts causing them to fail against manipulations that preserve audio-visual alignment. To address these limitations, we propose a single-stage training framework that enhances generalization by incorporating next-frame prediction for both uni-modal and cross-modal features. Additionally, we introduce a window-level attention mechanism to capture discrepancies between predicted and actual frames, enabling the model to detect local artifacts around every frame, which is crucial for accurately classifying fully manipulated videos and effectively localizing deepfake segments in partially spoofed samples. Our model, evaluated on multiple benchmark datasets, demonstrates strong generalization and precise temporal localization.

Yuchen Hu, Chen Chen, Chao-Han Huck Yang et al.

We propose an unsupervised adaptation framework, Self-TAught Recognizer (STAR), which leverages unlabeled data to enhance the robustness of automatic speech recognition (ASR) systems in diverse target domains, such as noise and accents. STAR is developed for prevalent speech foundation models based on Transformer-related architecture with auto-regressive decoding (e.g., Whisper, Canary). Specifically, we propose a novel indicator that empirically integrates step-wise information during decoding to assess the token-level quality of pseudo labels without ground truth, thereby guiding model updates for effective unsupervised adaptation. Experimental results show that STAR achieves an average of 13.5% relative reduction in word error rate across 14 target domains, and it sometimes even approaches the upper-bound performance of supervised adaptation. Surprisingly, we also observe that STAR prevents the adapted model from the common catastrophic forgetting problem without recalling source-domain data. Furthermore, STAR exhibits high data efficiency that only requires less than one-hour unlabeled data, and seamless generality to alternative large speech models and speech translation tasks. Our code aims to open source to the research communities.

Yizhou Peng, Yi-Wen Chao, Dianwen Ng et al.

Full-duplex spoken dialogue systems (FDSDS) enable more natural human-machine interactions by allowing real-time user interruptions and backchanneling, compared to traditional SDS that rely on turn-taking. However, existing benchmarks lack metrics for FD scenes, e.g., evaluating model performance during user interruptions. In this paper, we present a comprehensive FD benchmarking pipeline utilizing LLMs, TTS, and ASR to address this gap. It assesses FDSDS's ability to handle user interruptions, manage delays, and maintain robustness in challenging scenarios with diverse novel metrics. We applied our benchmark to three open-source FDSDS (Moshi, Freeze-omni, and VITA-1.5) using over 40 hours of generated speech, with 293 simulated conversations and 1,200 interruptions. The results show that all models continue to face challenges, such as failing to respond to user interruptions, under frequent disruptions and noisy conditions. Demonstrations, data, and code will be released.

Ziyang Ma, Yinghao Ma, Yanqiao Zhu et al.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.

Yuchen Hu, Chen Chen, Chao-Han Huck Yang et al.

Recent advances in large language models (LLMs) have stepped forward the development of multilingual speech and machine translation by its reduced representation errors and incorporated external knowledge. However, both translation tasks typically utilize beam search decoding and top-1 hypothesis selection for inference. These techniques struggle to fully exploit the rich information in the diverse N-best hypotheses, making them less optimal for translation tasks that require a single, high-quality output sequence. In this paper, we propose a new generative paradigm for translation tasks, namely "GenTranslate", which builds upon LLMs to generate better results from the diverse translation versions in N-best list. Leveraging the rich linguistic knowledge and strong reasoning abilities of LLMs, our new paradigm can integrate the rich information in N-best candidates to generate a higher-quality translation result. Furthermore, to support LLM finetuning, we build and release a HypoTranslate dataset that contains over 592K hypotheses-translation pairs in 11 languages. Experiments on various speech and machine translation benchmarks (e.g., FLEURS, CoVoST-2, WMT) demonstrate that our GenTranslate significantly outperforms the state-of-the-art model.

Chen Chen, Yuchen Hu, Siyin Wang et al.

An ideal multimodal agent should be aware of the quality of its input modalities. Recent advances have enabled large language models (LLMs) to incorporate auditory systems for handling various speech-related tasks. However, most audio LLMs remain unaware of the quality of the speech they process. This limitation arises because speech quality evaluation is typically excluded from multi-task training due to the lack of suitable datasets. To address this, we introduce the first natural language-based speech evaluation corpus, generated from authentic human ratings. In addition to the overall Mean Opinion Score (MOS), this corpus offers detailed analysis across multiple dimensions and identifies causes of quality degradation. It also enables descriptive comparisons between two speech samples (A/B tests) with human-like judgment. Leveraging this corpus, we propose an alignment approach with LLM distillation (ALLD) to guide the audio LLM in extracting relevant information from raw speech and generating meaningful responses. Experimental results demonstrate that ALLD outperforms the previous state-of-the-art regression model in MOS prediction, with a mean square error of 0.17 and an A/B test accuracy of 98.6%. Additionally, the generated responses achieve BLEU scores of 25.8 and 30.2 on two tasks, surpassing the capabilities of task-specific models. This work advances the comprehensive perception of speech signals by audio LLMs, contributing to the development of real-world auditory and sensory intelligent agents.

Yuchen Hu, Chen Chen, Chengwei Qin et al.

Recent advances in large language models (LLMs) have promoted generative error correction (GER) for automatic speech recognition (ASR), which aims to predict the ground-truth transcription from the decoded N-best hypotheses. Thanks to the strong language generation ability of LLMs and rich information in the N-best list, GER shows great effectiveness in enhancing ASR results. However, it still suffers from two limitations: 1) LLMs are unaware of the source speech during GER, which may lead to results that are grammatically correct but violate the source speech content, 2) N-best hypotheses usually only vary in a few tokens, making it redundant to send all of them for GER, which could confuse LLM about which tokens to focus on and thus lead to increased miscorrection. In this paper, we propose ClozeGER, a new paradigm for ASR generative error correction. First, we introduce a multimodal LLM (i.e., SpeechGPT) to receive source speech as extra input to improve the fidelity of correction output. Then, we reformat GER as a cloze test with logits calibration to remove the input information redundancy and simplify GER with clear instructions. Experiments show that ClozeGER achieves a new breakthrough over vanilla GER on 9 popular ASR datasets.

Xiangyu Zhang, Daijiao Liu, Hexin Liu et al.

Recently, Denoising Diffusion Probabilistic Models (DDPMs) have attained leading performances across a diverse range of generative tasks. However, in the field of speech synthesis, although DDPMs exhibit impressive performance, their long training duration and substantial inference costs hinder practical deployment. Existing approaches primarily focus on enhancing inference speed, while approaches to accelerate training a key factor in the costs associated with adding or customizing voices often necessitate complex modifications to the model, compromising their universal applicability. To address the aforementioned challenges, we propose an inquiry: is it possible to enhance the training/inference speed and performance of DDPMs by modifying the speech signal itself? In this paper, we double the training and inference speed of Speech DDPMs by simply redirecting the generative target to the wavelet domain. This method not only achieves comparable or superior performance to the original model in speech synthesis tasks but also demonstrates its versatility. By investigating and utilizing different wavelet bases, our approach proves effective not just in speech synthesis, but also in speech enhancement.

Haoyang Li, Jia Qi Yip, Tianyu Fan et al.

Recent advancements in Neural Audio Codec (NAC) models have inspired their use in various speech processing tasks, including speech enhancement (SE). In this work, we propose a novel, efficient SE approach by leveraging the pre-quantization output of a pretrained NAC encoder. Unlike prior NAC-based SE methods, which process discrete speech tokens using Language Models (LMs), we perform SE within the continuous embedding space of the pretrained NAC, which is highly compressed along the time dimension for efficient representation. Our lightweight SE model, optimized through an embedding-level loss, delivers results comparable to SE baselines trained on larger datasets, with a significantly lower real-time factor of 0.005. Additionally, our method achieves a low GMAC of 3.94, reducing complexity 18-fold compared to Sepformer in a simulated cloud-based audio transmission environment. This work highlights a new, efficient NAC-based SE solution, particularly suitable for cloud applications where NAC is used to compress audio before transmission. Copyright 20XX 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.

He Wang, Pengcheng Guo, Yue Li et al.

To promote speech processing and recognition research in driving scenarios, we build on the success of the Intelligent Cockpit Speech Recognition Challenge (ICSRC) held at ISCSLP 2022 and launch the ICASSP 2024 In-Car Multi-Channel Automatic Speech Recognition (ICMC-ASR) Challenge. This challenge collects over 100 hours of multi-channel speech data recorded inside a new energy vehicle and 40 hours of noise for data augmentation. Two tracks, including automatic speech recognition (ASR) and automatic speech diarization and recognition (ASDR) are set up, using character error rate (CER) and concatenated minimum permutation character error rate (cpCER) as evaluation metrics, respectively. Overall, the ICMC-ASR Challenge attracts 98 participating teams and receives 53 valid results in both tracks. In the end, first-place team USTCiflytek achieves a CER of 13.16% in the ASR track and a cpCER of 21.48% in the ASDR track, showing an absolute improvement of 13.08% and 51.4% compared to our challenge baseline, respectively.

Donghang Wu, Haoyang Zhang, Jun Chen et al.

Real-time Spoken Language Models (SLMs) struggle to leverage Chain-of-Thought (CoT) reasoning due to the prohibitive latency of generating the entire thought process sequentially. Enabling SLMs to think while speaking, similar to humans, is attracting increasing attention. We present, for the first time, Mind-Paced Speaking (MPS), a brain-inspired framework that enables high-fidelity, real-time reasoning. Similar to how humans utilize distinct brain regions for thinking and responding, we propose a novel dual-brain approach, employing a "Formulation Brain" for high-level reasoning to pace and guide a separate "Articulation Brain" for fluent speech generation. This division of labor eliminates mode-switching, preserving the integrity of the reasoning process. Experiments show that MPS significantly outperforms existing think-while-speaking methods and achieves reasoning performance comparable to models that pre-compute the full CoT before speaking, while drastically reducing latency. Under a zero-latency configuration, the proposed method achieves an accuracy of 92.8% on the mathematical reasoning task Spoken-MQA and attains a score of 82.5 on the speech conversation task URO-Bench. Our work effectively bridges the gap between high-quality reasoning and real-time interaction.

Donghang Wu, Haoyang Zhang, Chen Chen et al.

Recent advances in spoken dialogue language models (SDLMs) reflect growing interest in shifting from turn-based to full-duplex systems, where the models continuously perceive user speech streams while generating responses. This simultaneous listening and speaking design enables real-time interaction and the agent can handle dynamic conversational behaviors like user barge-in. However, during the listening phase, existing systems keep the agent idle by repeatedly predicting the silence token, which departs from human behavior: we usually engage in lightweight thinking during conversation rather than remaining absent-minded. Inspired by this, we propose Chronological Thinking, a on-the-fly conversational thinking mechanism that aims to improve response quality in full-duplex SDLMs. Specifically, chronological thinking presents a paradigm shift from conventional LLM thinking approaches, such as Chain-of-Thought, purpose-built for streaming acoustic input. (1) Strictly causal: the agent reasons incrementally while listening, updating internal hypotheses only from past audio with no lookahead. (2) No additional latency: reasoning is amortized during the listening window; once the user stops speaking, the agent halts thinking and begins speaking without further delay. Experiments demonstrate the effectiveness of chronological thinking through both objective metrics and human evaluations show consistent improvements in response quality. Furthermore, chronological thinking robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.

Fabian Ritter-Gutierrez, Yi-Cheng Lin, Jeremy H. M Wong et al.

Creating a unified speech and music model requires expensive pre-training. Model merging can instead create an unified audio model with minimal computational expense. However, direct merging is challenging when the models are not aligned in the weight space. Motivated by Git Re-Basin, we introduce a correlation-permutation approach that aligns a music encoder's internal layers with a speech encoder. We extend previous work to the case of merging transformer layers. The method computes a permutation matrix that maximizes the model's features-wise cross-correlations layer by layer, enabling effective fusion of these otherwise disjoint models. The merged model retains speech capabilities through this method while significantly enhancing music performance, achieving an improvement of 14.83 points in average score compared to linear interpolation model merging. This work allows the creation of unified audio models from independently trained encoders.

Alan Dao, Dinh Bach Vu, Huy Hoang Ha et al.

The rapid growth of voice assistants powered by large language models (LLM) has highlighted a need for speech instruction data to train these systems. Despite the abundance of speech recognition data, there is a notable scarcity of speech instruction data, which is essential for fine-tuning models to understand and execute spoken commands. Generating high-quality synthetic speech requires a good text-to-speech (TTS) model, which may not be available to low resource languages. Our novel approach addresses this challenge by halting synthesis at the semantic representation level, bypassing the need for TTS. We achieve this by aligning synthetic semantic representations with the pre-trained Whisper encoder, enabling an LLM to be fine-tuned on text instructions while maintaining the ability to understand spoken instructions during inference. This simplified training process is a promising approach to building voice assistant for low-resource languages.

Yizhou Peng, Bin Wang, Yi-Wen Chao et al.

This report details the NTU Speechlab system developed for the Interspeech 2025 Multilingual Conversational Speech and Language Model (MLC-SLM) Challenge (Task I), where we achieved 5th place. We present comprehensive analyses of our multilingual automatic speech recognition system, highlighting key advancements in model architecture, data selection, and training strategies. In particular, language-specific prompts and model averaging techniques were instrumental in boosting system performance across diverse languages. Compared to the initial baseline system, our final model reduced the average Mix Error Rate from 20.2% to 10.6%, representing an absolute improvement of 9.6% (a relative improvement of 48%) on the evaluation set. Our results demonstrate the effectiveness of our approach and offer practical insights for future Speech Large Language Models.

Yizhou Peng, Hexin Liu, Eng Siong Chng

This paper introduces the integration of language-specific bi-directional context into a speech large language model (SLLM) to improve multilingual continuous conversational automatic speech recognition (ASR). We propose a character-level contextual masking strategy during training, which randomly removes portions of the context to enhance robustness and better emulate the flawed transcriptions that may occur during inference. For decoding, a two-stage pipeline is utilized: initial isolated segment decoding followed by context-aware re-decoding using neighboring hypotheses. Evaluated on the 1500-hour Multilingual Conversational Speech and Language Model (MLC-SLM) corpus covering eleven languages, our method achieves an 18% relative improvement compared to a strong baseline, outperforming even the model trained on 6000 hours of data for the MLC-SLM competition. These results underscore the significant benefit of incorporating contextual information in multilingual continuous conversational ASR.

Heqing Zou, Fengmao Lv, Desheng Zheng et al.

Multilingual speech emotion recognition aims to estimate a speaker's emotional state using a contactless method across different languages. However, variability in voice characteristics and linguistic diversity poses significant challenges for zero-shot speech emotion recognition, especially with multilingual datasets. In this paper, we propose leveraging contrastive learning to refine multilingual speech features and extend large language models for zero-shot multilingual speech emotion estimation. Specifically, we employ a novel two-stage training framework to align speech signals with linguistic features in the emotional space, capturing both emotion-aware and language-agnostic speech representations. To advance research in this field, we introduce a large-scale synthetic multilingual speech emotion dataset, M5SER. Our experiments demonstrate the effectiveness of the proposed method in both speech emotion recognition and zero-shot multilingual speech emotion recognition, including previously unseen datasets and languages.

Haorui He, Yuchen Song, Yuancheng Wang et al.

The effectiveness of one-shot voice conversion (VC) decreases in real-world scenarios where reference speeches, which are often sourced from the internet, contain various disturbances like background noise. To address this issue, we introduce Noro, a noise-robust one-shot VC system. Noro features innovative components tailored for VC using noisy reference speeches, including a dual-branch reference encoding module and a noise-agnostic contrastive speaker loss. Experimental results demonstrate that Noro outperforms our baseline system in both clean and noisy scenarios, highlighting its efficacy for real-world applications. Additionally, we investigate the hidden speaker representation capabilities of our baseline system by repurposing its reference encoder as a speaker encoder. The results show that it is competitive with several advanced self-supervised learning models for speaker representation under the SUPERB settings, highlighting the potential for advancing speaker representation learning through one-shot VC tasks.

Yuxin Li, Eng Siong Chng, Cuntai Guan

Speech-based depression detection (SDD) is a promising, non-invasive alternative to traditional clinical assessments. However, it remains limited by the difficulty of extracting meaningful features and capturing sparse, heterogeneous depressive cues over time. Pretrained self-supervised learning (SSL) models such as WavLM provide rich, multi-layer speech representations, yet most existing SDD methods rely only on the final layer or search for a single best-performing one. These approaches often overfit to specific datasets and fail to leverage the full hierarchical structure needed to detect subtle and persistent depression signals. To address this challenge, we propose HAREN-CTC, a novel architecture that integrates multi-layer SSL features using cross-attention within a multitask learning framework, combined with Connectionist Temporal Classification loss to handle sparse temporal supervision. HAREN-CTC comprises two key modules: a Hierarchical Adaptive Clustering module that reorganizes SSL features into complementary embeddings, and a Cross-Modal Fusion module that models inter-layer dependencies through cross-attention. The CTC objective enables alignment-aware training, allowing the model to track irregular temporal patterns of depressive speech cues. We evaluate HAREN-CTC under both an upper-bound setting with standard data splits and a generalization setting using five-fold cross-validation. The model achieves state-of-the-art macro F1-scores of 0.81 on DAIC-WOZ and 0.82 on MODMA, outperforming prior methods across both evaluation scenarios.

Duc-Tuan Truong, Tianchi Liu, Ruijie Tao et al.

Recent work shows that one-class learning can detect unseen deepfake attacks by modeling a compact distribution of bona fide speech around a single centroid. However, the single-centroid assumption can oversimplify the bona fide speech representation and overlook useful cues, such as speech quality, which reflects the naturalness of the speech. Speech quality can be easily obtained using existing speech quality assessment models that estimate it through Mean Opinion Score. In this paper, we propose QAMO: Quality-Aware Multi-Centroid One-Class Learning for speech deepfake detection. QAMO extends conventional one-class learning by introducing multiple quality-aware centroids. In QAMO, each centroid is optimized to represent a distinct speech quality subspaces, enabling better modeling of intra-class variability in bona fide speech. In addition, QAMO supports a multi-centroid ensemble scoring strategy, which improves decision thresholding and reduces the need for quality labels during inference. With two centroids to represent high- and low-quality speech, our proposed QAMO achieves an equal error rate of 5.09% in In-the-Wild dataset, outperforming previous one-class and quality-aware systems.

Chin Yuen Kwok, Jia Qi Yip, Eng Siong Chng

Rare word recognition can be improved by adapting ASR models to synthetic data that includes these words. Further improvements can be achieved through contextual biasing, which trains and adds a biasing module into the model architecture to prioritize rare words. While training the module on synthetic rare word data is more effective than using non-rare-word data, it can lead to overfitting due to artifacts in the synthetic audio. To address this, we enhance the TCPGen-based contextual biasing approach and propose a keyword-aware loss function that additionally focuses on biased words when training biasing modules. This loss includes a masked cross-entropy term for biased word prediction and a binary classification term for detecting biased word positions. These two terms complementarily support the decoding of biased words during inference. By adapting Whisper to 10 hours of synthetic data, our method reduced the word error rate on the NSC Part 2 test set from 29.71% to 11.81%.

Sathya Krishnan Suresh, Tanmay Surana, Lim Zhi Hao et al.

Code-switching (CS) poses a significant challenge for Large Language Models (LLMs), yet its comprehensibility remains underexplored in LLMs. We introduce CS-Sum, to evaluate the comprehensibility of CS by the LLMs through CS dialogue to English summarization. CS-Sum is the first benchmark for CS dialogue summarization across Mandarin-English (EN-ZH), Tamil-English (EN-TA), and Malay-English (EN-MS), with 900-1300 human-annotated dialogues per language pair. Evaluating ten LLMs, including open and closed-source models, we analyze performance across few-shot, translate-summarize, and fine-tuning (LoRA, QLoRA on synthetic data) approaches. Our findings show that though the scores on automated metrics are high, LLMs make subtle mistakes that alter the complete meaning of the dialogue. To this end, we introduce 3 most common type of errors that LLMs make when handling CS input. Error rates vary across CS pairs and LLMs, with some LLMs showing more frequent errors on certain language pairs, underscoring the need for specialized training on code-switched data.