Zhuo Chen

Sanyuan Chen, Yu Wu, Chengyi Wang et al. · microsoft-research

The massive growth of self-supervised learning (SSL) has been witnessed in language, vision, speech, and audio domains over the past few years. While discrete label prediction is widely adopted for other modalities, the state-of-the-art audio SSL models still employ reconstruction loss for pre-training. Compared with reconstruction loss, semantic-rich discrete label prediction encourages the SSL model to abstract the high-level audio semantics and discard the redundant details as in human perception. However, a semantic-rich acoustic tokenizer for general audio pre-training is usually not straightforward to obtain, due to the continuous property of audio and unavailable phoneme sequences like speech. To tackle this challenge, we propose BEATs, an iterative audio pre-training framework to learn Bidirectional Encoder representation from Audio Transformers, where an acoustic tokenizer and an audio SSL model are optimized by iterations. In the first iteration, we use random projection as the acoustic tokenizer to train an audio SSL model in a mask and label prediction manner. Then, we train an acoustic tokenizer for the next iteration by distilling the semantic knowledge from the pre-trained or fine-tuned audio SSL model. The iteration is repeated with the hope of mutual promotion of the acoustic tokenizer and audio SSL model. The experimental results demonstrate our acoustic tokenizers can generate discrete labels with rich audio semantics and our audio SSL models achieve state-of-the-art results across various audio classification benchmarks, even outperforming previous models that use more training data and model parameters significantly. Specifically, we set a new state-of-the-art mAP 50.6% on AudioSet-2M for audio-only models without using any external data, and 98.1% accuracy on ESC-50. The code and pre-trained models are available at https://aka.ms/beats.

Naoyuki Kanda, Jian Wu, Xiaofei Wang et al.

This paper presents a novel streaming automatic speech recognition (ASR) framework for multi-talker overlapping speech captured by a distant microphone array with an arbitrary geometry. Our framework, named t-SOT-VA, capitalizes on independently developed two recent technologies; array-geometry-agnostic continuous speech separation, or VarArray, and streaming multi-talker ASR based on token-level serialized output training (t-SOT). To combine the best of both technologies, we newly design a t-SOT-based ASR model that generates a serialized multi-talker transcription based on two separated speech signals from VarArray. We also propose a pre-training scheme for such an ASR model where we simulate VarArray's output signals based on monaural single-talker ASR training data. Conversation transcription experiments using the AMI meeting corpus show that the system based on the proposed framework significantly outperforms conventional ones. Our system achieves the state-of-the-art word error rates of 13.7% and 15.5% for the AMI development and evaluation sets, respectively, in the multiple-distant-microphone setting while retaining the streaming inference capability.

Xueyao Zhang, Chaoren Wang, Huan Liao et al.

Aligning large generative models with human feedback is a critical challenge. In speech synthesis, this is particularly pronounced due to the lack of a large-scale human preference dataset, which hinders the development of models that truly align with human perception. To address this, we introduce SpeechJudge, a comprehensive suite comprising a dataset, a benchmark, and a reward model centered on naturalness--one of the most fundamental subjective metrics for speech synthesis. First, we present SpeechJudge-Data, a large-scale human feedback corpus of 99K speech pairs. The dataset is constructed using a diverse set of advanced zero-shot text-to-speech (TTS) models across diverse speech styles and multiple languages, with human annotations for both intelligibility and naturalness preference. From this, we establish SpeechJudge-Eval, a challenging benchmark for speech naturalness judgment. Our evaluation reveals that existing metrics and AudioLLMs struggle with this task; the leading model, Gemini-2.5-Flash, achieves less than 70% agreement with human judgment, highlighting a significant gap for improvement. To bridge this gap, we develop SpeechJudge-GRM, a generative reward model (GRM) based on Qwen2.5-Omni-7B. It is trained on SpeechJudge-Data via a two-stage post-training process: Supervised Fine-Tuning (SFT) with Chain-of-Thought rationales followed by Reinforcement Learning (RL) with GRPO on challenging cases. On the SpeechJudge-Eval benchmark, the proposed SpeechJudge-GRM demonstrates superior performance, achieving 77.2% accuracy (and 79.4% after inference-time scaling @10) compared to a classic Bradley-Terry reward model (72.7%). Furthermore, SpeechJudge-GRM can be also employed as a reward function during the post-training of speech generation models to facilitate their alignment with human preferences.

Chenda Li, Jing Shi, Wangyou Zhang et al.

We present ESPnet-SE, which is designed for the quick development of speech enhancement and speech separation systems in a single framework, along with the optional downstream speech recognition module. ESPnet-SE is a new project which integrates rich automatic speech recognition related models, resources and systems to support and validate the proposed front-end implementation (i.e. speech enhancement and separation).It is capable of processing both single-channel and multi-channel data, with various functionalities including dereverberation, denoising and source separation. We provide all-in-one recipes including data pre-processing, feature extraction, training and evaluation pipelines for a wide range of benchmark datasets. This paper describes the design of the toolkit, several important functionalities, especially the speech recognition integration, which differentiates ESPnet-SE from other open source toolkits, and experimental results with major benchmark datasets.

Yi Yuan, Zhuo Chen, Xubo Liu et al.

Contrastive language-audio pretraining~(CLAP) has been developed to align the representations of audio and language, achieving remarkable performance in retrieval and classification tasks. However, current CLAP struggles to capture temporal information within audio and text features, presenting substantial limitations for tasks such as audio retrieval and generation. To address this gap, we introduce T-CLAP, a temporal-enhanced CLAP model. We use Large Language Models~(LLMs) and mixed-up strategies to generate temporal-contrastive captions for audio clips from extensive audio-text datasets. Subsequently, a new temporal-focused contrastive loss is designed to fine-tune the CLAP model by incorporating these synthetic data. We conduct comprehensive experiments and analysis in multiple downstream tasks. T-CLAP shows improved capability in capturing the temporal relationship of sound events and outperforms state-of-the-art models by a significant margin.

Xueyao Zhang, Junan Zhang, Yuancheng Wang et al.

Controllable human voice generation, particularly for expressive domains like singing, remains a significant challenge. This paper introduces Vevo2, a unified framework for controllable speech and singing voice generation. To tackle issues like the scarcity of annotated singing data and to enable flexible controllability, Vevo2 introduces two audio tokenizers: (1) a music-notation-free prosody tokenizer that captures prosody and melody from speech, singing, and even instrumental sounds, and (2) a low-frame-rate (12.5 Hz) content-style tokenizer that encodes linguistic content, prosody, and style for both speech and singing, while enabling timbre disentanglement. Vevo2 consists of an auto-regressive (AR) content-style modeling stage, which aims to enable controllability over text, prosody, and style, as well as a flow-matching acoustic modeling stage that allows for timbre control. Particularly, during pre-training of the AR model, we propose both explicit and implicit prosody learning strategies to bridge speech and singing voice. Moreover, to further enhance the AR model's ability to follow text and prosody, we design a multi-objective post-training task that integrates both intelligibility and prosody similarity alignment. Experimental results show that the unified modeling in Vevo2 brings mutual benefits to both speech and singing voice generation. Additionally, Vevo2's effectiveness across a wide range of synthesis, conversion, and editing tasks for both speech and singing further demonstrates its strong generalization ability and versatility. Audio samples are are available at https://versasinger.github.io/.

Wangyou Zhang, Zhuo Chen, Naoyuki Kanda et al.

Multi-talker conversational speech processing has drawn many interests for various applications such as meeting transcription. Speech separation is often required to handle overlapped speech that is commonly observed in conversation. Although the original utterancelevel permutation invariant training-based continuous speech separation approach has proven to be effective in various conditions, it lacks the ability to leverage the long-span relationship of utterances and is computationally inefficient due to the highly overlapped sliding windows. To overcome these drawbacks, we propose a novel training scheme named Group-PIT, which allows direct training of the speech separation models on the long-form speech with a low computational cost for label assignment. Two different speech separation approaches with Group-PIT are explored, including direct long-span speech separation and short-span speech separation with long-span tracking. The experiments on the simulated meeting-style data demonstrate the effectiveness of our proposed approaches, especially in dealing with a very long speech input.

Desh Raj, Liang Lu, Zhuo Chen et al.

Streaming recognition of multi-talker conversations has so far been evaluated only for 2-speaker single-turn sessions. In this paper, we investigate it for multi-turn meetings containing multiple speakers using the Streaming Unmixing and Recognition Transducer (SURT) model, and show that naively extending the single-turn model to this harder setting incurs a performance penalty. As a solution, we propose the dual-path (DP) modeling strategy first used for time-domain speech separation. We experiment with LSTM and Transformer based DP models, and show that they improve word error rate (WER) performance while yielding faster convergence. We also explore training strategies such as chunk width randomization and curriculum learning for these models, and demonstrate their importance through ablation studies. Finally, we evaluate our models on the LibriCSS meeting data, where they perform competitively with offline separation-based methods.

Yi Luo, Zhuo Chen, Cong Han et al.

Modules in all existing speech separation networks can be categorized into single-input-multi-output (SIMO) modules and single-input-single-output (SISO) modules. SIMO modules generate more outputs than input, and SISO modules keep the numbers of input and output the same. While the majority of separation models only contain SIMO architectures, it has also been shown that certain two-stage separation systems integrated with a post-enhancement SISO module can improve the separation quality. Why performance improvements can be achieved by incorporating the SISO modules? Are SIMO modules always necessary? In this paper, we empirically examine those questions by designing models with varying configurations in the SIMO and SISO modules. We show that comparing with the standard SIMO-only design, a mixed SIMO-SISO design with a same model size is able to improve the separation performance especially under low-overlap conditions. We further validate the necessity of SIMO modules and show that SISO-only models are still able to perform separation without sacrificing the performance. The observations allow us to rethink the model design paradigm and present different views on how the separation is performed.

Sanyuan Chen, Yu Wu, Zhuo Chen et al.

Continuous speech separation plays a vital role in complicated speech related tasks such as conversation transcription. The separation model extracts a single speaker signal from a mixed speech. In this paper, we use transformer and conformer in lieu of recurrent neural networks in the separation system, as we believe capturing global information with the self-attention based method is crucial for the speech separation. Evaluating on the LibriCSS dataset, the conformer separation model achieves state of the art results, with a relative 23.5% word error rate (WER) reduction from bi-directional LSTM (BLSTM) in the utterance-wise evaluation and a 15.4% WER reduction in the continuous evaluation.

Zhuo Chen, Takuya Yoshioka, Liang Lu et al.

This paper describes a dataset and protocols for evaluating continuous speech separation algorithms. Most prior studies on speech separation use pre-segmented signals of artificially mixed speech utterances which are mostly \emph{fully} overlapped, and the algorithms are evaluated based on signal-to-distortion ratio or similar performance metrics. However, in natural conversations, a speech signal is continuous, containing both overlapped and overlap-free components. In addition, the signal-based metrics have very weak correlations with automatic speech recognition (ASR) accuracy. We think that not only does this make it hard to assess the practical relevance of the tested algorithms, it also hinders researchers from developing systems that can be readily applied to real scenarios. In this paper, we define continuous speech separation (CSS) as a task of generating a set of non-overlapped speech signals from a \textit{continuous} audio stream that contains multiple utterances that are \emph{partially} overlapped by a varying degree. A new real recorded dataset, called LibriCSS, is derived from LibriSpeech by concatenating the corpus utterances to simulate a conversation and capturing the audio replays with far-field microphones. A Kaldi-based ASR evaluation protocol is also established by using a well-trained multi-conditional acoustic model. By using this dataset, several aspects of a recently proposed speaker-independent CSS algorithm are investigated. The dataset and evaluation scripts are available to facilitate the research in this direction.

Yi Luo, Zhuo Chen, Nima Mesgarani et al.

An important problem in ad-hoc microphone speech separation is how to guarantee the robustness of a system with respect to the locations and numbers of microphones. The former requires the system to be invariant to different indexing of the microphones with the same locations, while the latter requires the system to be able to process inputs with varying dimensions. Conventional optimization-based beamforming techniques satisfy these requirements by definition, while for deep learning-based end-to-end systems those constraints are not fully addressed. In this paper, we propose transform-average-concatenate (TAC), a simple design paradigm for channel permutation and number invariant multi-channel speech separation. Based on the filter-and-sum network (FaSNet), a recently proposed end-to-end time-domain beamforming system, we show how TAC significantly improves the separation performance across various numbers of microphones in noisy reverberant separation tasks with ad-hoc arrays. Moreover, we show that TAC also significantly improves the separation performance with fixed geometry array configuration, further proving the effectiveness of the proposed paradigm in the general problem of multi-microphone speech separation.

Zhuo Chen, Yi Luo, Nima Mesgarani

Despite the overwhelming success of deep learning in various speech processing tasks, the problem of separating simultaneous speakers in a mixture remains challenging. Two major difficulties in such systems are the arbitrary source permutation and unknown number of sources in the mixture. We propose a novel deep learning framework for single channel speech separation by creating attractor points in high dimensional embedding space of the acoustic signals which pull together the time-frequency bins corresponding to each source. Attractor points in this study are created by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. The proposed model is different from prior works in that it implements an end-to-end training, and it does not depend on the number of sources in the mixture. Two strategies are explored in the test time, K-means and fixed attractor points, where the latter requires no post-processing and can be implemented in real-time. We evaluated our system on Wall Street Journal dataset and show 5.49\% improvement over the previous state-of-the-art methods.