Chen Chen

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.

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.

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).

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.

Chenyang Song, Xu Han, Zhengyan Zhang et al.

Activation sparsity refers to the existence of considerable weakly-contributed elements among activation outputs. As a prevalent property of the models using the ReLU activation function, activation sparsity has been proven a promising paradigm to boost model inference efficiency. Nevertheless, most large language models (LLMs) adopt activation functions without intrinsic activation sparsity (e.g., GELU and Swish). Some recent efforts have explored introducing ReLU or its variants as the substitutive activation function to help LLMs achieve activation sparsity and inference acceleration, but few can simultaneously obtain high sparsity and comparable model performance. This paper introduces a simple and effective sparsification method named "ProSparse" to push LLMs for higher activation sparsity while maintaining comparable performance. Specifically, after substituting the activation function of LLMs with ReLU, ProSparse adopts progressive sparsity regularization with a factor smoothly increasing along the multi-stage sine curves. This can enhance activation sparsity and mitigate performance degradation by avoiding radical shifts in activation distributions. With ProSparse, we obtain high sparsity of 89.32% for LLaMA2-7B, 88.80% for LLaMA2-13B, and 87.89% for end-size MiniCPM-1B, respectively, achieving comparable performance to their original Swish-activated versions. These present the most sparsely activated models among open-source LLaMA versions and competitive end-size models, considerably surpassing ReluLLaMA-7B (66.98%) and ReluLLaMA-13B (71.56%). Our inference acceleration experiments further demonstrate the significant practical acceleration potential of LLMs with higher activation sparsity, obtaining up to 4.52$\times$ inference speedup.

Shengji Jin, Yuanhao Zou, Victor Zhu et al.

While Multimodal Large Language Models (MLLMs) have advanced Video Temporal Grounding (VTG), existing methods often couple output paradigms with different backbones, datasets, and training protocols. This makes it challenging to isolate the specific impact of the output design. Additionally, as VTG systems are increasingly considered for resource-constrained edge deployment, the trade-off between output formulation and system-level efficiency requires systematic investigation. In this paper, we present a controlled empirical study comparing three dominant VTG output paradigms: Text Numeral Generation, Temporal Token Generation, and Continuous Temporal Decoding. We evaluate these paradigms across identical compact VLMs (SmolVLM2, FastVLM, and Molmo2) using consistent datasets and LoRA fine-tuning protocols. Evaluations on Charades-STA, QVHighlights, and YouCook2 measure both localization accuracy and system efficiency, including inference latency, training throughput, and parameter overhead. Our results demonstrate that the choice of output formulation significantly affects both grounding accuracy and computational cost, independent of model scale. Specifically, the continuous distribution paradigm consistently achieves the most favorable efficiency-accuracy trade-off on the Pareto frontier, delivering robust localization with minimal latency overhead. These findings provide objective empirical guidelines for designing efficient, deployment-ready VTG systems.

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.

Yingfa Chen, Zhengyan Zhang, Xu Han et al. · tsinghua

Large language models (LLMs) can make predictions using parametric knowledge--knowledge encoded in the model weights--or contextual knowledge--knowledge presented in the context. In many scenarios, a desirable behavior is that LLMs give precedence to contextual knowledge when it conflicts with the parametric knowledge, and fall back to using their parametric knowledge when the context is irrelevant. This enables updating and correcting the model's knowledge by in-context editing instead of retraining. Previous works have shown that LLMs are inclined to ignore contextual knowledge and fail to reliably fall back to parametric knowledge when presented with irrelevant context. In this work, we discover that, with proper prompting methods, instruction-finetuned LLMs can be highly controllable by contextual knowledge and robust to irrelevant context. Utilizing this feature, we propose EREN (Edit models by REading Notes) to improve the scalability and robustness of LLM editing. To better evaluate the robustness of model editors, we collect a new dataset, that contains irrelevant questions that are more challenging than the ones in existing datasets. Empirical results show that our method outperforms current state-of-the-art methods by a large margin. Unlike existing techniques, it can integrate knowledge from multiple edits, and correctly respond to syntactically similar but semantically unrelated inputs (and vice versa). The source code can be found at https://github.com/thunlp/EREN.

Qian Yong, Chen Chen, Xiabing Zhou

Contrastive learning has been frequently investigated to learn effective representations for text clustering tasks. While existing contrastive learning-based text clustering methods only focus on modeling instance-wise semantic similarity relationships, they ignore contextual information and underlying relationships among all instances that needs to be clustered. In this paper, we propose a novel text clustering approach called Subspace Contrastive Learning (SCL) which models cluster-wise relationships among instances. Specifically, the proposed SCL consists of two main modules: (1) a self-expressive module that constructs virtual positive samples and (2) a contrastive learning module that further learns a discriminative subspace to capture task-specific cluster-wise relationships among texts. Experimental results show that the proposed SCL method not only has achieved superior results on multiple task clustering datasets but also has less complexity in positive sample construction.

Chen Chen, Ruizhe Li, Yuchen Hu et al. · gatech

Recent studies have successfully shown that large language models (LLMs) can be successfully used for generative error correction (GER) on top of the automatic speech recognition (ASR) output. Specifically, an LLM is utilized to carry out a direct mapping from the N-best hypotheses list generated by an ASR system to the predicted output transcription. However, despite its effectiveness, GER introduces extra data uncertainty since the LLM is trained without taking into account acoustic information available in the speech signal. In this work, we aim to overcome such a limitation by infusing acoustic information before generating the predicted transcription through a novel late fusion solution termed Uncertainty-Aware Dynamic Fusion (UADF). UADF is a multimodal fusion approach implemented into an auto-regressive decoding process and works in two stages: (i) It first analyzes and calibrates the token-level LLM decision, and (ii) it then dynamically assimilates the information from the acoustic modality. Experimental evidence collected from various ASR tasks shows that UADF surpasses existing fusion mechanisms in several ways. It yields significant improvements in word error rate (WER) while mitigating data uncertainty issues in LLM and addressing the poor generalization relied with sole modality during fusion. We also demonstrate that UADF seamlessly adapts to audio-visual speech recognition.

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.

Ke Hu, Ehsan Hosseini-Asl, Chen Chen et al. · nvidia

Spoken dialogue is an intuitive form of human-computer interaction, yet current speech language models often remain constrained to turn-based exchanges, lacking real-time adaptability such as user barge-in. We propose a novel duplex speech to speech (S2S) architecture featuring continuous user inputs and codec agent outputs with channel fusion that directly models simultaneous user and agent streams. Using a pretrained streaming encoder for user input enables the first duplex S2S model without requiring speech pretrain. Separate architectures for agent and user modeling facilitate codec fine-tuning for better agent voices and halve the bitrate (0.6 kbps) compared to previous works. Experimental results show that the proposed model outperforms previous duplex models in reasoning, turn-taking, and barge-in abilities. The model requires significantly less speech data, as speech pretrain is skipped, which markedly simplifies the process of building a duplex S2S model from any LLMs. Finally, it is the first openly available duplex S2S model with training and inference code to foster reproducibility.

Mengyuan Liu, Chen Chen, Songtao Wu et al.

Recognizing interactive actions, including hand-to-hand interaction and human-to-human interaction, has attracted increasing attention for various applications in the field of video analysis and human-robot interaction. Considering the success of graph convolution in modeling topology-aware features from skeleton data, recent methods commonly operate graph convolution on separate entities and use late fusion for interactive action recognition, which can barely model the mutual semantic relationships between pairwise entities. To this end, we propose a mutual excitation graph convolutional network (me-GCN) by stacking mutual excitation graph convolution (me-GC) layers. Specifically, me-GC uses a mutual topology excitation module to firstly extract adjacency matrices from individual entities and then adaptively model the mutual constraints between them. Moreover, me-GC extends the above idea and further uses a mutual feature excitation module to extract and merge deep features from pairwise entities. Compared with graph convolution, our proposed me-GC gradually learns mutual information in each layer and each stage of graph convolution operations. Extensive experiments on a challenging hand-to-hand interaction dataset, i.e., the Assembely101 dataset, and two large-scale human-to-human interaction datasets, i.e., NTU60-Interaction and NTU120-Interaction consistently verify the superiority of our proposed method, which outperforms the state-of-the-art GCN-based and Transformer-based methods.

Md Kowsher, Nusrat Jahan Prottasha, Shiyun Xu et al.

Bidirectional language models have better context understanding and perform better than unidirectional models on natural language understanding tasks, yet the theoretical reasons behind this advantage remain unclear. In this work, we investigate this disparity through the lens of the Information Bottleneck (IB) principle, which formalizes a trade-off between compressing input information and preserving task-relevant content. We propose FlowNIB, a dynamic and scalable method for estimating mutual information during training that addresses key limitations of classical IB approaches, including computational intractability and fixed trade-off schedules. Theoretically, we show that bidirectional models retain more mutual information and exhibit higher effective dimensionality than unidirectional models. To support this, we present a generalized framework for measuring representational complexity and prove that bidirectional representations are strictly more informative under mild conditions. We further validate our findings through extensive experiments across multiple models and tasks using FlowNIB, revealing how information is encoded and compressed throughout training. Together, our work provides a principled explanation for the effectiveness of bidirectional architectures and introduces a practical tool for analyzing information flow in deep language models.

Chen Chen, Zehua Liu, Xiaolou Li et al.

The first Chinese Continuous Visual Speech Recognition Challenge aimed to probe the performance of Large Vocabulary Continuous Visual Speech Recognition (LVC-VSR) on two tasks: (1) Single-speaker VSR for a particular speaker and (2) Multi-speaker VSR for a set of registered speakers. The challenge yielded highly successful results, with the best submission significantly outperforming the baseline, particularly in the single-speaker task. This paper comprehensively reviews the challenge, encompassing the data profile, task specifications, and baseline system construction. It also summarises the representative techniques employed by the submitted systems, highlighting the most effective approaches. Additional information and resources about this challenge can be accessed through the official website at http://cnceleb.org/competition.

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

Recent advances in large language models (LLMs) have promoted generative error correction (GER) for automatic speech recognition (ASR), which leverages the rich linguistic knowledge and powerful reasoning ability of LLMs to improve recognition results. The latest work proposes a GER benchmark with HyPoradise dataset to learn the mapping from ASR N-best hypotheses to ground-truth transcription by efficient LLM finetuning, which shows great effectiveness but lacks specificity on noise-robust ASR. In this work, we extend the benchmark to noisy conditions and investigate if we can teach LLMs to perform denoising for GER just like what robust ASR do}, where one solution is introducing noise information as a conditioner into LLM. However, directly incorporating noise embeddings from audio encoder could harm the LLM tuning due to cross-modality gap. To this end, we propose to extract a language-space noise embedding from the N-best list to represent the noise conditions of source speech, which can promote the denoising process in GER. Furthermore, in order to enhance its representation ability of audio noise, we design a knowledge distillation (KD) approach via mutual information estimation to distill the real noise information in audio embeddings to our language embedding. Experiments on various latest LLMs demonstrate our approach achieves a new breakthrough with up to 53.9% correction improvement in terms of word error rate while with limited training data. Analysis shows that our language-space noise embedding can well represent the noise conditions of source speech, under which off-the-shelf LLMs show strong ability of language-space denoising.

Chen Chen, Chao-Han Huck Yang, Kai Li et al.

In this work, we introduce S4M, a new efficient speech separation framework based on neural state-space models (SSM). Motivated by linear time-invariant systems for sequence modeling, our SSM-based approach can efficiently model input signals into a format of linear ordinary differential equations (ODEs) for representation learning. To extend the SSM technique into speech separation tasks, we first decompose the input mixture into multi-scale representations with different resolutions. This mechanism enables S4M to learn globally coherent separation and reconstruction. The experimental results show that S4M performs comparably to other separation backbones in terms of SI-SDRi, while having a much lower model complexity with significantly fewer trainable parameters. In addition, our S4M-tiny model (1.8M parameters) even surpasses attention-based Sepformer (26.0M parameters) in noisy conditions with only 9.2 of multiply-accumulate operation (MACs).