In this paper, we focus on prompting one of the most important tasks in the field of speech processing, i.e., automatic speech recognition (ASR), with speech foundation encoders and large language models (LLM). Despite the growing body of research in this area, we find that many crucial design decisions in LLM-based ASR systems are often inadequately justified. This lack of clarity impedes the field's progress, making it challenging to pinpoint which design choices truly improve model performance. To address these challenges, we conduct a comprehensive series of experiments that explore various aspects, leading to the optimal LLM-based ASR system. We found that delicate designs are not necessary, while a clean setup with little task-specific design is competent. The models achieve strong performance on the Librispeech and Gigaspeech datasets, compared to both LLM-based models and non-LLM-based models. Finally, we explore the capability emergence of LLM-based ASR in the process of modal alignment. We hope that our study can facilitate the research on extending LLM with cross-modality capacity and shed light on the LLM-based ASR community.

Qinglin Zhang, Luyao Cheng, Chong Deng, Qian Chen, Wen Wang, Siqi Zheng, Jiaqing Liu, Hai Yu, Chao-Hong Tan, Zhihao Du, ShiLiang Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Efficient numerical representation is crucial for deep learning accelerators, especially for large language models (LLMs). The 8-bit-floating-point (FP8) data representation achieves higher precision and fewer quantization efforts than integer, which has been proven inevitable in attention-based accelerators for LLMs. Therefore, area-efficient design techniques for FP8 play a central role in lowering LLMs chip’s budget. This paper presents AttenPU, which is built upon reconfigurable FP8 units and supports E4M3 for inference and E5M2 for training. Bidirectional dataflow is exploited to enable AttenPU to interact with FP32 coprocessor to reduce latency. The design achieves a low FP8-to-INT8 area ratio of 1.63×, an area efficiency of 193.5 GFLOPS/mm2, with an 87.05% reduction in the latency of RTX3090 GPU.

LLM-based text-to-speech(TTS) system has becoming the new trend and SOTA due to its high naturalness and zero-shot capability. However, it relies heavily on training data, usually requires at least thousands hours of labeled audio. In this report, we describe how to use pretrained CosyVoice model, to develop a streaming TTS system which supports Indian English and Indian languages. Though the pretrained CosyVoice model has never seen such data, it shows good performance in both specific speaker TTS and zero-shot voice clone after finetuning with merely 280 hours data. Experiment on LIMMITS25 challenge shows that our system achieves 4.46/4.19/4.55 naturalness, and 4.29/4.34/4.27 similarity in track1/track2/track3 respectively, which ranked 1st in all tracks.

While automatic speech recognition (ASR) systems have achieved remarkable performance with large-scale datasets, their efficacy remains inadequate in low-resource settings, encompassing dialects, accents, minority languages, and long-tail hotwords, domains with significant practical relevance. With the advent of versatile and powerful text-to-speech (TTS) models, capable of generating speech with human-level naturalness, expressiveness, and diverse speaker profiles, leveraging TTS for ASR data augmentation provides a cost-effective and practical approach to enhancing ASR performance. Comprehensive experiments on an unprecedentedly rich variety of low-resource datasets demonstrate consistent and substantial performance improvements, proving that the proposed method of enhancing low-resource ASR through a versatile TTS model is highly effective and has broad application prospects. Furthermore, we delve deeper into key characteristics of synthesized speech data that contribute to ASR improvement, examining factors such as text diversity, speaker diversity, and the volume of synthesized data, with text diversity being studied for the first time in this work. We hope our findings provide helpful guidance and reference for the practical application of TTS-based data augmentation and push the advancement of low-resource ASR one step further.

Human speech goes beyond the mere transfer of information; it is a profound exchange of emotions and a connection between individuals. While Text-to-Speech (TTS) models have made huge progress, they still face challenges in controlling the emotional expression in the generated speech. In this work, we propose EmoVoice, a novel emotion-controllable TTS model that exploits large language models (LLMs) to enable fine-grained freestyle natural language emotion control, and a phoneme boost variant design that makes the model output phoneme tokens and audio tokens in parallel to enhance content consistency, inspired by chain-of-thought (CoT) and modality-of-thought (CoM) techniques. Besides, we introduce EmoVoice-DB, a high-quality 40-hour English emotion dataset featuring expressive speech and fine-grained emotion labels with natural language descriptions. EmoVoice achieves state-of-the-art performance on the English EmoVoice-DB test set using only synthetic training data, and on the Chinese Secap test set using our in-house data. We further investigate the reliability of existing emotion evaluation metrics and their alignment with human perceptual preferences, and explore using SOTA multimodal LLMs GPT-4o-audio and Gemini to assess emotional speech. Dataset, code, checkpoints and demo samples are available at https://github.com/yanghaha0908/EmoVoice.

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.

This paper presents FunCodec, a fundamental neural speech codec toolkit, which is an extension of the open-source speech processing toolkit FunASR. FunCodec provides reproducible training recipes and inference scripts for the latest neural speech codec models, such as SoundStream and Encodec. Thanks to the unified design with FunASR, FunCodec can be easily integrated into downstream tasks, such as speech recognition. Along with FunCodec, pretrained models are also provided, which can be used for academic or generalized purposes. Based on the toolkit, we further propose the frequency-domain codec models, FreqCodec, which can achieve comparable speech quality with much lower computation and parameter complexity. Experimental results show that, under the same compression ratio, FunCodec can achieve better reconstruction quality compared with other toolkits and released models. We also demonstrate that the pre-trained models are suitable for downstream tasks, including automatic speech recognition and personalized text-to-speech synthesis. This toolkit is publicly available at https://github.com/alibaba-damo-academy/FunCodec.

Joint modeling of multi-speaker ASR and speaker diarization has recently shown promising results in speaker-attributed automatic speech recognition (SA-ASR). Although being able to obtain state-of-the-art (SOTA) performance, most of the studies are based on an autoregressive (AR) decoder which generates tokens one-by-one and results in a large real-time factor (RTF). To speed up inference, we introduce a recently proposed non-autoregressive model Paraformer as an acoustic model in the SA-ASR model. Paraformer uses a single-step decoder to enable parallel generation, obtaining comparable performance to the SOTA AR transformer models. Besides, we propose a speaker-filling strategy to reduce speaker identification errors and adopt an inter-CTC strategy to enhance the encoder’s ability in acoustic modeling. Experiments on the AliMeeting corpus show that our model outperforms the cascaded SA-ASR model by a 6.1% relative speaker-dependent character error rate (SD-CER) reduction on the test set. Moreover, our model achieves a comparable SD-CER of 34.8% with only 1/10 RTF compared with the SOTA joint AR SA-ASR model.

With the success of the first Multi-channel Multi-party Meeting Transcription challenge (M2MeT), the second M2MeT challenge (M2MeT 2.0) held in ASRU2023 particularly aims to tackle the complex task of speaker-attributed ASR (SAASR), which directly addresses the practical and challenging problem of “who spoke what at when” at typical meeting scenario. We particularly established two sub-tracks. The fixed training condition sub-track, where the training data is constrained to predetermined datasets, but participants can use any open-source pre-trained model. The open training condition sub-track, which allows for the use of all available data and models without limitation. In addition, we release a new 10-hour test set for challenge ranking. This paper provides an overview of the dataset, track settings, results, and analysis of submitted systems, as a benchmark to show the current state of speaker-attributed ASR.

Recently, end-to-end neural diarization (EEND) is introduced and achieves promising results in speaker-overlapped scenarios. In EEND, speaker diarization is formulated as a multi-label prediction problem, where speaker activities are estimated independently and their dependency are not well considered. To overcome these disadvantages, we employ the power set encoding to reformulate speaker diarization as a single-label classification problem and propose the overlap-aware EEND (EEND-OLA) model, in which speaker overlaps and dependency can be modeled explicitly. Inspired by the success of two-stage hybrid systems, we further propose a novel Two-stage OverLap-aware Diarization framework (TOLD) by involving a speaker overlap-aware post-processing (SOAP) model to iteratively refine the diarization results of EEND-OLA. Experimental results show that, compared with the original EEND, the proposed EEND-OLA achieves a 14.39% relative improvement in terms of diarization error rates (DER), and utilizing SOAP provides another 19.33% relative improvement. As a result, our method TOLD achieves a DER of 10.14% on the CALLHOME dataset, which is a new state-of-the-art result on this benchmark to the best of our knowledge.

Recently, hybrid systems of clustering and neural diarization models have been successfully applied in multi-party meeting analysis.However, current models always treat overlapped speaker diarization as a multi-label classification problem, where speaker dependency and overlaps are not well considered.To overcome the disadvantages, we reformulate overlapped speaker diarization task as a single-label prediction problem via the proposed power set encoding (PSE).Through this formulation, speaker dependency and overlaps can be explicitly modeled.To fully leverage this formulation, we further propose the speaker overlap-aware neural diarization (SOND) model, which consists of a contextindependent (CI) scorer to model global speaker discriminability, a context-dependent scorer (CD) to model local discriminability, and a speaker combining network (SCN) to combine and reassign speaker activities.Experimental results show that using the proposed formulation can outperform the state-ofthe-art methods based on target speaker voice activity detection, and the performance can be further improved with SOND, resulting in a 6.30% relative diarization error reduction.

Recent development of speech signal processing, such as speech recognition, speaker diarization, etc., has inspired numerous applications of speech technologies. The meeting scenario is one of the most valuable and, at the same time, most challenging scenarios for the deployment of speech technologies. Speaker diarization and multi-speaker automatic speech recognition in meeting scenarios have attracted much attention recently. However, the lack of large public meeting data has been a major obstacle for advancement of the field. Therefore, we make available the AliMeeting corpus, which consists of 120 hours of recorded Mandarin meeting data, including far-field data collected by 8-channel microphone array as well as near-field data collected by headset microphone. Each meeting session is composed of 2-4 speakers with different speaker overlap ratio, recorded in meeting rooms with different size. Along with the dataset, we launch the ICASSP 2022 Multi-channel Multi-party Meeting Transcription Challenge (M2MeT) with two tracks, namely speaker diarization and multi-speaker ASR, aiming to provide a common testbed for meeting rich transcription and promote reproducible research in this field. In this paper we provide a detailed introduction of the AliMeeting dateset, challenge rules, evaluation methods and baseline systems.

The ICASSP 2022 Multi-channel Multi-party Meeting Transcription Grand Challenge (M2MeT) focuses on one of the most valuable and the most challenging scenarios of speech technologies. The M2MeT challenge has particularly set up two tracks, speaker diarization (track 1) and multi-speaker automatic speech recognition (ASR) (track 2). Along with the challenge, we released 120 hours of real-recorded Mandarin meeting speech data with manual annotation, including far-field data collected by 8-channel micro-phone array as well as near-field data collected by each participants’ headset microphone. We briefly describe the released dataset, track setups, baselines and summarize the challenge results and major techniques used in the submissions.

In this paper, we conduct a comparative study on speaker-attributed automatic speech recognition (SA-ASR) in the multi-party meeting scenario, a topic with increasing attention in meeting rich transcription. Specifically, three approaches are evaluated in this study. The first approach, FD-SOT, consists of a frame-level diarization model to identify speakers and a multi-talker ASR to recognize utterances. The speaker-attributed transcriptions are obtained by aligning the diarization results and recognized hypotheses. However, such an alignment strategy may suffer from erroneous timestamps due to the modular independence, severely hindering the model performance. Therefore, we propose the second approach, WD-SOT, to address alignment errors by introducing a word-level diarization model, which can get rid of such timestamp alignment dependency. To further mitigate the alignment issues, we propose the third approach, TS-ASR, which trains a target-speaker separation module and an ASR module jointly. By comparing various strategies for each SA-ASR approach, experimental results on a real meeting scenario corpus, AliMeeting, reveal that the WD-SOT approach achieves 10.7% relative reduction on averaged speaker-dependent character error rate (SD-CER), compared with the FD-SOT approach. In addition, the TS-ASR approach also outperforms the FD-SOT approach and brings 16.5% relative average SD-CER reduction.

Recently cross-channel attention, which better leverages multi-channel signals from microphone array, has shown promising results in the multi-party meeting scenario. Cross-channel attention focuses on either learning global correlations between sequences of different channels or exploiting fine-grained channel-wise information effectively at each time step. Considering the delay of microphone array receiving sound, we propose a multi-frame cross-channel attention, which models cross-channel information between adjacent frames to exploit the complementarity of both frame-wise and channel-wise knowledge. Besides, we also propose a multi-layer convolutional mechanism to fuse the multi -channel output and a channel masking strategy to combat the channel number mismatch problem between training and inference. Experiments on the AliMeeting, a real-world corpus, reveal that our proposed model outperforms single-channel model by 31.7% and 37.0% CER reduction on Eval and Test sets. Moreover, with comparable model parameters and training data, our proposed model achieves a new SOTA performance on the AliMeeting corpus, as compared with the top ranking systems in the ICASSP2022 M2MeT challenge, a recently held multi-channel multi-speaker ASR challenge.

This paper focuses on the problem of temporal dependency modeling in the CNN-based models for audio classification tasks. To capture audio temporal dependencies using CNNs, we take a different approach from the purely architecture-induced method and explicitly encode temporal dependencies into the CNN-based audio classifiers. More specifically, in addition to the classification objective, we require the CNN model to solve an auxiliary task of predicting the future features, which is formulated by leveraging the Contrastive Predictive Coding (CPC) loss. Furthermore, a novel hierarchical CPC (HCPC) model is proposed for capturing multi-level temporal dependencies at the same time. The proposed model is evaluated on a wide range of non-speech audio signals, including musical and in-the-wild environmental audio signals. We show that the proposed approach improves the backbone CNNs consistently on all tested benchmark datasets and outperforms a DenseNet model trained from scratch.

Current methods for monaural speech enhancement only utilize acoustic information but seldom consider the phonetic information of an utterance. In the voice conversion community, significant progress has been achieved by using the phonetic information via the phonetic posteriorgrams (PPGs). Inspired by the progress, we propose a phoneme-aware network (PAN) to utilize the noisy PPGs for speech enhancement. Since the PPG prediction and speech enhancement benefit from each other, a PPG predictor is involved into the PAN and an iterative training algorithm is proposed for PAN. Experimental results show that the enhancement performance is improved by using the phonetic information in terms of speech intelligibility, perceptual quality and character error rate. To the best of our knowledge, this is the first time to introduce the PPG into speech enhancement.

Conventional monaural speech enhancement methods usually enhance the magnitude spectrum of noisy speech and leave the phase unchanged. Recent studies suggest that phase is also important for both speech intelligibility and perceptual quality. Although deep learning exhibits great potential on enhancing the magnitude and phase spectra in complex spectrogram domain and waveform domain, complex spectrogram and waveform are always more difficult to predict than the magnitude spectrum due to lack of clear structure in them. In this study, a Mel-domain denoising autoencoder and a deep generative vocoder are stacked to form a joint framework for monaural speech enhancement, in which the clean speech waveform is reconstructed without using the phase. Specifically, a convolutional recurrent network (CRN) is employed as the denoising autoencoder to enhance the Mel power spectrum of noisy speech. Then, the enhanced Mel power spectrum is fed to a deep generative vocoder to synthesize the speech waveform. Furthermore, the denoising autoencoder and generative vocoder are jointly fine-tuned. Experimental results show that the proposed method significantly improves speech intelligibility and perceptual quality. More importantly, our method achieves much better generalization ability for untrained noises than previous methods.

In this paper, we propose a new strategy for acoustic scene classification (ASC) , namely recognizing acoustic scenes through identifying distinct sound events. This differs from existing strategies, which focus on characterizing global acoustical distributions of audio or the temporal evolution of short-term audio features, without analysis down to the level of sound events. To identify distinct sound events for each scene, we formulate ASC in a multi-instance learning (MIL) framework, where each audio recording is mapped into a bag-of-instances representation. Here, instances can be seen as high-level representations for sound events inside a scene. We also propose a MIL neural networks model, which implicitly identifies distinct instances (i.e., sound events). Furthermore, we propose two specially designed modules that model the multi-temporal scale and multi-modal natures of the sound events respectively. The experiments were conducted on the official development set of the DCASE2018 Task1 Subtask B, and our best-performing model improves over the official baseline by 9.4% (68.3% vs 58.9%) in terms of classification accuracy. This study indicates that recognizing acoustic scenes by identifying distinct sound events is effective and paves the way for future studies that combine this strategy with previous ones.