Zhijie Yan

Keyu An, Qian Chen, Chong Deng et al.

This report introduces FunAudioLLM, a model family designed to enhance natural voice interactions between humans and large language models (LLMs). At its core are two innovative models: SenseVoice, which handles multilingual speech recognition, emotion recognition, and audio event detection; and CosyVoice, which facilitates natural speech generation with control over multiple languages, timbre, speaking style, and speaker identity. SenseVoice-Small delivers exceptionally low-latency ASR for 5 languages, and SenseVoice-Large supports high-precision ASR for over 50 languages, while CosyVoice excels in multi-lingual voice generation, zero-shot in-context learning, cross-lingual voice cloning, and instruction-following capabilities. The models related to SenseVoice and CosyVoice have been open-sourced on Modelscope and Huggingface, along with the corresponding training, inference, and fine-tuning codes released on GitHub. By integrating these models with LLMs, FunAudioLLM enables applications such as speech-to-speech translation, emotional voice chat, interactive podcasts, and expressive audiobook narration, thereby pushing the boundaries of voice interaction technology. Demos are available at https://fun-audio-llm.github.io, and the code can be accessed at https://github.com/FunAudioLLM.

Zhifu Gao, Shiliang Zhang, Ian McLoughlin et al.

Transformers have recently dominated the ASR field. Although able to yield good performance, they involve an autoregressive (AR) decoder to generate tokens one by one, which is computationally inefficient. To speed up inference, non-autoregressive (NAR) methods, e.g. single-step NAR, were designed, to enable parallel generation. However, due to an independence assumption within the output tokens, performance of single-step NAR is inferior to that of AR models, especially with a large-scale corpus. There are two challenges to improving single-step NAR: Firstly to accurately predict the number of output tokens and extract hidden variables; secondly, to enhance modeling of interdependence between output tokens. To tackle both challenges, we propose a fast and accurate parallel transformer, termed Paraformer. This utilizes a continuous integrate-and-fire based predictor to predict the number of tokens and generate hidden variables. A glancing language model (GLM) sampler then generates semantic embeddings to enhance the NAR decoder's ability to model context interdependence. Finally, we design a strategy to generate negative samples for minimum word error rate training to further improve performance. Experiments using the public AISHELL-1, AISHELL-2 benchmark, and an industrial-level 20,000 hour task demonstrate that the proposed Paraformer can attain comparable performance to the state-of-the-art AR transformer, with more than 10x speedup.

Yunfei Chu, Jin Xu, Xiaohuan Zhou et al.

Recently, instruction-following audio-language models have received broad attention for audio interaction with humans. However, the absence of pre-trained audio models capable of handling diverse audio types and tasks has hindered progress in this field. Consequently, most existing works have only been able to support a limited range of interaction capabilities. In this paper, we develop the Qwen-Audio model and address this limitation by scaling up audio-language pre-training to cover over 30 tasks and various audio types, such as human speech, natural sounds, music, and songs, to facilitate universal audio understanding abilities. However, directly co-training all tasks and datasets can lead to interference issues, as the textual labels associated with different datasets exhibit considerable variations due to differences in task focus, language, granularity of annotation, and text structure. To overcome the one-to-many interference, we carefully design a multi-task training framework by conditioning on a sequence of hierarchical tags to the decoder for encouraging knowledge sharing and avoiding interference through shared and specified tags respectively. Remarkably, Qwen-Audio achieves impressive performance across diverse benchmark tasks without requiring any task-specific fine-tuning, surpassing its counterparts. Building upon the capabilities of Qwen-Audio, we further develop Qwen-Audio-Chat, which allows for input from various audios and text inputs, enabling multi-turn dialogues and supporting various audio-central scenarios.

Zhihao Du, Qian Chen, Shiliang Zhang et al.

Recent years have witnessed a trend that large language model (LLM) based text-to-speech (TTS) emerges into the mainstream due to their high naturalness and zero-shot capacity. In this paradigm, speech signals are discretized into token sequences, which are modeled by an LLM with text as prompts and reconstructed by a token-based vocoder to waveforms. Obviously, speech tokens play a critical role in LLM-based TTS models. Current speech tokens are learned in an unsupervised manner, which lacks explicit semantic information and alignment to the text. In this paper, we propose to represent speech with supervised semantic tokens, which are derived from a multilingual speech recognition model by inserting vector quantization into the encoder. Based on the tokens, we further propose a scalable zero-shot TTS synthesizer, CosyVoice, which consists of an LLM for text-to-token generation and a conditional flow matching model for token-to-speech synthesis. Experimental results show that supervised semantic tokens significantly outperform existing unsupervised tokens in terms of content consistency and speaker similarity for zero-shot voice cloning. Moreover, we find that utilizing large-scale data further improves the synthesis performance, indicating the scalable capacity of CosyVoice. To the best of our knowledge, this is the first attempt to involve supervised speech tokens into TTS models.

Zhihao Du, Jiaming Wang, Qian Chen et al.

Generative Pre-trained Transformer (GPT) models have achieved remarkable performance on various natural language processing tasks, and have shown great potential as backbones for audio-and-text large language models (LLMs). Previous mainstream audio-and-text LLMs use discrete audio tokens to represent both input and output audio; however, they suffer from performance degradation on tasks such as automatic speech recognition, speech-to-text translation, and speech enhancement over models using continuous speech features. In this paper, we propose LauraGPT, a novel unified audio-and-text GPT-based LLM for audio recognition, understanding, and generation. LauraGPT is a versatile LLM that can process both audio and text inputs and generate outputs in either modalities. We propose a novel data representation that combines continuous and discrete features for audio: LauraGPT encodes input audio into continuous representations using an audio encoder and generates output audio from discrete codec codes. We propose a one-step codec vocoder to overcome the prediction challenge caused by the multimodal distribution of codec tokens. We fine-tune LauraGPT using supervised multi-task learning. Extensive experiments show that LauraGPT consistently achieves comparable to superior performance compared to strong baselines on a wide range of audio tasks related to content, semantics, paralinguistics, and audio-signal analysis, such as automatic speech recognition, speech-to-text translation, text-to-speech synthesis, speech enhancement, automated audio captioning, speech emotion recognition, and spoken language understanding.

Xiaohuan Zhou, Jiaming Wang, Zeyu Cui et al.

In this paper, we propose a novel multi-modal multi-task encoder-decoder pre-training framework (MMSpeech) for Mandarin automatic speech recognition (ASR), which employs both unlabeled speech and text data. The main difficulty in speech-text joint pre-training comes from the significant difference between speech and text modalities, especially for Mandarin speech and text. Unlike English and other languages with an alphabetic writing system, Mandarin uses an ideographic writing system where character and sound are not tightly mapped to one another. Therefore, we propose to introduce the phoneme modality into pre-training, which can help capture modality-invariant information between Mandarin speech and text. Specifically, we employ a multi-task learning framework including five self-supervised and supervised tasks with speech and text data. For end-to-end pre-training, we introduce self-supervised speech-to-pseudo-codes (S2C) and phoneme-to-text (P2T) tasks utilizing unlabeled speech and text data, where speech-pseudo-codes pairs and phoneme-text pairs are a supplement to the supervised speech-text pairs. To train the encoder to learn better speech representation, we introduce self-supervised masked speech prediction (MSP) and supervised phoneme prediction (PP) tasks to learn to map speech into phonemes. Besides, we directly add the downstream supervised speech-to-text (S2T) task into the pre-training process, which can further improve the pre-training performance and achieve better recognition results even without fine-tuning. Experiments on AISHELL-1 show that our proposed method achieves state-of-the-art performance, with a more than 40% relative improvement compared with other pre-training methods.

Xian Shi, Yanni Chen, Shiliang Zhang et al.

Conventional ASR systems use frame-level phoneme posterior to conduct force-alignment~(FA) and provide timestamps, while end-to-end ASR systems especially AED based ones are short of such ability. This paper proposes to perform timestamp prediction~(TP) while recognizing by utilizing continuous integrate-and-fire~(CIF) mechanism in non-autoregressive ASR model - Paraformer. Foucing on the fire place bias issue of CIF, we conduct post-processing strategies including fire-delay and silence insertion. Besides, we propose to use scaled-CIF to smooth the weights of CIF output, which is proved beneficial for both ASR and TP task. Accumulated averaging shift~(AAS) and diarization error rate~(DER) are adopted to measure the quality of timestamps and we compare these metrics of proposed system and conventional hybrid force-alignment system. The experiment results over manually-marked timestamps testset show that the proposed optimization methods significantly improve the accuracy of CIF timestamps, reducing 66.7\% and 82.1\% of AAS and DER respectively. Comparing to Kaldi force-alignment trained with the same data, optimized CIF timestamps achieved 12.3\% relative AAS reduction.

Zhijie Yan, Shufei Li, Ze Zhang et al.

Reliable mobile manipulation in dynamic indoor environments requires a scene representation that remains geometrically consistent, semantically queryable, and computationally bounded as the environment changes. Existing systems often rely on pre-built maps, static-scene assumptions, or highly accurate camera poses, which can lead to stale or misaligned scene information when target objects are relocated or pose estimates are corrected. This paper presents DREAM, a real-robot mobile manipulation framework that integrates perception, memory, localization, navigation, and manipulation in previously unseen indoor environments without a pre-built map. DREAM constructs an online spatio-semantic voxel memory from RGB-D observations registered by a LiDAR-inertial-visual SLAM backend. It further introduces pose-graph-aware Redundancy-Aware Memory Pruning (RMP) to update historical observations after pose corrections while keeping long-horizon observation history bounded. For target localization and reacquisition, DREAM combines language-conditioned 3D retrieval, open-vocabulary image detection, and multimodal large language model based semantic verification. Real-robot experiments in four dynamic indoor laboratory scenes show that DREAM improves long-horizon task success rates from 40%-60% with DynaMem to 55%-70%, while maintaining a memory footprint of 0.37-0.63 GB and an online memory-update time of 0.43-0.53 s across scenes.

Qinglin Zhang, Chong Deng, Jiaqing Liu et al.

Listening to long video/audio recordings from video conferencing and online courses for acquiring information is extremely inefficient. Even after ASR systems transcribe recordings into long-form spoken language documents, reading ASR transcripts only partly speeds up seeking information. It has been observed that a range of NLP applications, such as keyphrase extraction, topic segmentation, and summarization, significantly improve users' efficiency in grasping important information. The meeting scenario is among the most valuable scenarios for deploying these spoken language processing (SLP) capabilities. However, the lack of large-scale public meeting datasets annotated for these SLP tasks severely hinders their advancement. To prompt SLP advancement, we establish a large-scale general Meeting Understanding and Generation Benchmark (MUG) to benchmark the performance of a wide range of SLP tasks, including topic segmentation, topic-level and session-level extractive summarization and topic title generation, keyphrase extraction, and action item detection. To facilitate the MUG benchmark, we construct and release a large-scale meeting dataset for comprehensive long-form SLP development, the AliMeeting4MUG Corpus, which consists of 654 recorded Mandarin meeting sessions with diverse topic coverage, with manual annotations for SLP tasks on manual transcripts of meeting recordings. To the best of our knowledge, the AliMeeting4MUG Corpus is so far the largest meeting corpus in scale and facilitates most SLP tasks. In this paper, we provide a detailed introduction of this corpus, SLP tasks and evaluation methods, baseline systems and their performance.

Zhihao Du, Shiliang Zhang, Siqi Zheng et al.

Overlapping speech diarization has been traditionally treated as a multi-label classification problem. In this paper, we reformulate this task as a single-label prediction problem by encoding multiple binary labels into a single label with the power set, which represents the possible combinations of target speakers. This formulation has two benefits. First, the overlaps of target speakers are explicitly modeled. Second, threshold selection is no longer needed. Through this formulation, we propose the speaker embedding-aware neural diarization (SEND) framework, where a speech encoder, a speaker encoder, two similarity scorers, and a post-processing network are jointly optimized to predict the encoded labels according to the similarities between speech features and speaker embeddings. Experimental results show that SEND has a stable learning process and can be trained on highly overlapped data without extra initialization. More importantly, our method achieves the state-of-the-art performance in real meeting scenarios with fewer model parameters and lower computational complexity.

Qinglin Zhang, Chong Deng, Jiaqing Liu et al.

ICASSP2023 General Meeting Understanding and Generation Challenge (MUG) focuses on prompting a wide range of spoken language processing (SLP) research on meeting transcripts, as SLP applications are critical to improve users' efficiency in grasping important information in meetings. MUG includes five tracks, including topic segmentation, topic-level and session-level extractive summarization, topic title generation, keyphrase extraction, and action item detection. To facilitate MUG, we construct and release a large-scale meeting dataset, the AliMeeting4MUG Corpus.

Keyu An, Shiliang Zhang, Zhijie Yan

In this study, we delve into the efficacy of transformers within pre-trained language models (PLMs) when repurposed as encoders for Automatic Speech Recognition (ASR). Our underlying hypothesis posits that, despite being initially trained on text-based corpora, these transformers possess a remarkable capacity to extract effective features from the input sequence. This inherent capability, we argue, is transferrable to speech data, thereby augmenting the acoustic modeling ability of ASR. Through rigorous empirical analysis, our findings reveal a notable improvement in Character Error Rate (CER) and Word Error Rate (WER) across diverse ASR tasks when transformers from pre-trained LMs are incorporated. Particularly, they serve as an advantageous starting point for initializing ASR encoders. Furthermore, we uncover that these transformers, when integrated into a well-established ASR encoder, can significantly boost performance, especially in scenarios where profound semantic comprehension is pivotal. This underscores the potential of leveraging the semantic prowess embedded within pre-trained transformers to advance ASR systems' capabilities.

Bu Jin, Yupeng Zheng, Pengfei Li et al.

3D dense captioning stands as a cornerstone in achieving a comprehensive understanding of 3D scenes through natural language. It has recently witnessed remarkable achievements, particularly in indoor settings. However, the exploration of 3D dense captioning in outdoor scenes is hindered by two major challenges: 1) the domain gap between indoor and outdoor scenes, such as dynamics and sparse visual inputs, makes it difficult to directly adapt existing indoor methods; 2) the lack of data with comprehensive box-caption pair annotations specifically tailored for outdoor scenes. To this end, we introduce the new task of outdoor 3D dense captioning. As input, we assume a LiDAR point cloud and a set of RGB images captured by the panoramic camera rig. The expected output is a set of object boxes with captions. To tackle this task, we propose the TOD3Cap network, which leverages the BEV representation to generate object box proposals and integrates Relation Q-Former with LLaMA-Adapter to generate rich captions for these objects. We also introduce the TOD3Cap dataset, the largest one to our knowledge for 3D dense captioning in outdoor scenes, which contains 2.3M descriptions of 64.3K outdoor objects from 850 scenes. Notably, our TOD3Cap network can effectively localize and caption 3D objects in outdoor scenes, which outperforms baseline methods by a significant margin (+9.6 CiDEr@0.5IoU). Code, data, and models are publicly available at https://github.com/jxbbb/TOD3Cap.

Xiaoji Zheng, Lixiu Wu, Zhijie Yan et al.

Motion prediction is among the most fundamental tasks in autonomous driving. Traditional methods of motion forecasting primarily encode vector information of maps and historical trajectory data of traffic participants, lacking a comprehensive understanding of overall traffic semantics, which in turn affects the performance of prediction tasks. In this paper, we utilized Large Language Models (LLMs) to enhance the global traffic context understanding for motion prediction tasks. We first conducted systematic prompt engineering, visualizing complex traffic environments and historical trajectory information of traffic participants into image prompts -- Transportation Context Map (TC-Map), accompanied by corresponding text prompts. Through this approach, we obtained rich traffic context information from the LLM. By integrating this information into the motion prediction model, we demonstrate that such context can enhance the accuracy of motion predictions. Furthermore, considering the cost associated with LLMs, we propose a cost-effective deployment strategy: enhancing the accuracy of motion prediction tasks at scale with 0.7\% LLM-augmented datasets. Our research offers valuable insights into enhancing the understanding of traffic scenes of LLMs and the motion prediction performance of autonomous driving. The source code is available at \url{https://github.com/AIR-DISCOVER/LLM-Augmented-MTR} and \url{https://aistudio.baidu.com/projectdetail/7809548}.

Huadai Liu, Tianyi Luo, Kaicheng Luo et al.

Traditional video-to-audio generation techniques primarily focus on perspective video and non-spatial audio, often missing the spatial cues necessary for accurately representing sound sources in 3D environments. To address this limitation, we introduce a novel task, 360V2SA, to generate spatial audio from 360-degree videos, specifically producing First-order Ambisonics (FOA) audio - a standard format for representing 3D spatial audio that captures sound directionality and enables realistic 3D audio reproduction. We first create Sphere360, a novel dataset tailored for this task that is curated from real-world data. We also design an efficient semi-automated pipeline for collecting and cleaning paired video-audio data. To generate spatial audio from 360-degree video, we propose a novel framework OmniAudio, which leverages self-supervised pre-training using both spatial audio data (in FOA format) and large-scale non-spatial data. Furthermore, OmniAudio features a dual-branch framework that utilizes both panoramic and perspective video inputs to capture comprehensive local and global information from 360-degree videos. Experimental results demonstrate that OmniAudio achieves state-of-the-art performance across both objective and subjective metrics on Sphere360. Code and datasets are available at https://github.com/liuhuadai/OmniAudio. The project website is available at https://OmniAudio-360V2SA.github.io.

Zhihao Du, Yuxuan Wang, Qian Chen et al.

In our previous work, we introduced CosyVoice, a multilingual speech synthesis model based on supervised discrete speech tokens. By employing progressive semantic decoding with two popular generative models, language models (LMs) and Flow Matching, CosyVoice demonstrated high prosody naturalness, content consistency, and speaker similarity in speech in-context learning. Recently, significant progress has been made in multi-modal large language models (LLMs), where the response latency and real-time factor of speech synthesis play a crucial role in the interactive experience. Therefore, in this report, we present an improved streaming speech synthesis model, CosyVoice 2, which incorporates comprehensive and systematic optimizations. Specifically, we introduce finite-scalar quantization to improve the codebook utilization of speech tokens. For the text-speech LM, we streamline the model architecture to allow direct use of a pre-trained LLM as the backbone. In addition, we develop a chunk-aware causal flow matching model to support various synthesis scenarios, enabling both streaming and non-streaming synthesis within a single model. By training on a large-scale multilingual dataset, CosyVoice 2 achieves human-parity naturalness, minimal response latency, and virtually lossless synthesis quality in the streaming mode. We invite readers to listen to the demos at https://funaudiollm.github.io/cosyvoice2.

Qian Chen, Yafeng Chen, Yanni Chen et al.

Recent advancements in large language models (LLMs) and multimodal speech-text models have laid the groundwork for seamless voice interactions, enabling real-time, natural, and human-like conversations. Previous models for voice interactions are categorized as native and aligned. Native models integrate speech and text processing in one framework but struggle with issues like differing sequence lengths and insufficient pre-training. Aligned models maintain text LLM capabilities but are often limited by small datasets and a narrow focus on speech tasks. In this work, we introduce MinMo, a Multimodal Large Language Model with approximately 8B parameters for seamless voice interaction. We address the main limitations of prior aligned multimodal models. We train MinMo through multiple stages of speech-to-text alignment, text-to-speech alignment, speech-to-speech alignment, and duplex interaction alignment, on 1.4 million hours of diverse speech data and a broad range of speech tasks. After the multi-stage training, MinMo achieves state-of-the-art performance across various benchmarks for voice comprehension and generation while maintaining the capabilities of text LLMs, and also facilitates full-duplex conversation, that is, simultaneous two-way communication between the user and the system. Moreover, we propose a novel and simple voice decoder that outperforms prior models in voice generation. The enhanced instruction-following capabilities of MinMo supports controlling speech generation based on user instructions, with various nuances including emotions, dialects, and speaking rates, and mimicking specific voices. For MinMo, the speech-to-text latency is approximately 100ms, full-duplex latency is approximately 600ms in theory and 800ms in practice. The MinMo project web page is https://funaudiollm.github.io/minmo, and the code and models will be released soon.

Zhengyan Sheng, Zhihao Du, Shiliang Zhang et al.

This paper presents a dual-stream text-to-speech (TTS) model, SyncSpeech, capable of receiving streaming text input from upstream models while simultaneously generating streaming speech, facilitating seamless interaction with large language models. SyncSpeech has the following advantages: Low latency, as it begins generating streaming speech upon receiving the second text token; High efficiency, as it decodes all speech tokens corresponding to the each arrived text token in one step. To achieve this, we propose a temporal masked transformer as the backbone of SyncSpeech, combined with token-level duration prediction to predict speech tokens and the duration for the next step. Additionally, we design a two-stage training strategy to improve training efficiency and the quality of generated speech. We evaluated the SyncSpeech on both English and Mandarin datasets. Compared to the recent dual-stream TTS models, SyncSpeech significantly reduces the first packet delay of speech tokens and accelerates the real-time factor. Moreover, with the same data scale, SyncSpeech achieves performance comparable to that of traditional autoregressive-based TTS models in terms of both speech quality and robustness. Speech samples are available at https://SyncSpeech.github.io/}{https://SyncSpeech.github.io/.

Xian Shi, Haoneng Luo, Zhifu Gao et al.

Estimating confidence scores for recognition results is a classic task in ASR field and of vital importance for kinds of downstream tasks and training strategies. Previous end-to-end~(E2E) based confidence estimation models (CEM) predict score sequences of equal length with input transcriptions, leading to unreliable estimation when deletion and insertion errors occur. In this paper we proposed CIF-Aligned confidence estimation model (CA-CEM) to achieve accurate and reliable confidence estimation based on novel non-autoregressive E2E ASR model - Paraformer. CA-CEM utilizes the modeling character of continuous integrate-and-fire (CIF) mechanism to generate token-synchronous acoustic embedding, which solves the estimation failure issue above. We measure the quality of estimation with AUC and RMSE in token level and ECE-U - a proposed metrics in utterance level. CA-CEM gains 24% and 19% relative reduction on ECE-U and also better AUC and RMSE on two test sets. Furthermore, we conduct analysis to explore the potential of CEM for different ASR related usage.

Yi Ren, Ming Lei, Zhiying Huang et al.

Expressive text-to-speech (TTS) has become a hot research topic recently, mainly focusing on modeling prosody in speech. Prosody modeling has several challenges: 1) the extracted pitch used in previous prosody modeling works have inevitable errors, which hurts the prosody modeling; 2) different attributes of prosody (e.g., pitch, duration and energy) are dependent on each other and produce the natural prosody together; and 3) due to high variability of prosody and the limited amount of high-quality data for TTS training, the distribution of prosody cannot be fully shaped. To tackle these issues, we propose ProsoSpeech, which enhances the prosody using quantized latent vectors pre-trained on large-scale unpaired and low-quality text and speech data. Specifically, we first introduce a word-level prosody encoder, which quantizes the low-frequency band of the speech and compresses prosody attributes in the latent prosody vector (LPV). Then we introduce an LPV predictor, which predicts LPV given word sequence. We pre-train the LPV predictor on large-scale text and low-quality speech data and fine-tune it on the high-quality TTS dataset. Finally, our model can generate expressive speech conditioned on the predicted LPV. Experimental results show that ProsoSpeech can generate speech with richer prosody compared with baseline methods.

Fan Yu, Shiliang Zhang, Pengcheng Guo et al.

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 microphone 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.

Fan Yu, Shiliang Zhang, Yihui Fu et al.

Recent development of speech 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. Specifically, two typical tasks, speaker diarization and multi-speaker automatic speech recognition have attracted much attention recently. However, the lack of large public meeting data has been a major obstacle for the 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 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.

Siqi Zheng, Shiliang Zhang, Weilong Huang et al.

We propose BeamTransformer, an efficient architecture to leverage beamformer's edge in spatial filtering and transformer's capability in context sequence modeling. BeamTransformer seeks to optimize modeling of sequential relationship among signals from different spatial direction. Overlapping speech detection is one of the tasks where such optimization is favorable. In this paper we effectively apply BeamTransformer to detect overlapping segments. Comparing to single-channel approach, BeamTransformer exceeds in learning to identify the relationship among different beam sequences and hence able to make predictions not only from the acoustic signals but also the localization of the source. The results indicate that a successful incorporation of microphone array signals can lead to remarkable gains. Moreover, BeamTransformer takes one step further, as speech from overlapped speakers have been internally separated into different beams.

Siqi Zheng, Weilong Huang, Xianliang Wang et al.

In this paper we describe a speaker diarization system that enables localization and identification of all speakers present in a conversation or meeting. We propose a novel systematic approach to tackle several long-standing challenges in speaker diarization tasks: (1) to segment and separate overlapping speech from two speakers; (2) to estimate the number of speakers when participants may enter or leave the conversation at any time; (3) to provide accurate speaker identification on short text-independent utterances; (4) to track down speakers movement during the conversation; (5) to detect speaker change incidence real-time. First, a differential directional microphone array-based approach is exploited to capture the target speakers' voice in far-field adverse environment. Second, an online speaker-location joint clustering approach is proposed to keep track of speaker location. Third, an instant speaker number detector is developed to trigger the mechanism that separates overlapped speech. The results suggest that our system effectively incorporates spatial information and achieves significant gains.

Shiliang Zhang, Zhifu Gao, Haoneng Luo et al.

Recently, streaming end-to-end automatic speech recognition (E2E-ASR) has gained more and more attention. Many efforts have been paid to turn the non-streaming attention-based E2E-ASR system into streaming architecture. In this work, we propose a novel online E2E-ASR system by using Streaming Chunk-Aware Multihead Attention(SCAMA) and a latency control memory equipped self-attention network (LC-SAN-M). LC-SAN-M uses chunk-level input to control the latency of encoder. As to SCAMA, a jointly trained predictor is used to control the output of encoder when feeding to decoder, which enables decoder to generate output in streaming manner. Experimental results on the open 170-hour AISHELL-1 and an industrial-level 20000-hour Mandarin speech recognition tasks show that our approach can significantly outperform the MoChA-based baseline system under comparable setup. On the AISHELL-1 task, our proposed method achieves a character error rate (CER) of 7.39%, to the best of our knowledge, which is the best published performance for online ASR.

Kai Fan, Jiayi Wang, Bo Li et al.

The performances of automatic speech recognition (ASR) systems are usually evaluated by the metric word error rate (WER) when the manually transcribed data are provided, which are, however, expensively available in the real scenario. In addition, the empirical distribution of WER for most ASR systems usually tends to put a significant mass near zero, making it difficult to simulate with a single continuous distribution. In order to address the two issues of ASR quality estimation (QE), we propose a novel neural zero-inflated model to predict the WER of the ASR result without transcripts. We design a neural zero-inflated beta regression on top of a bidirectional transformer language model conditional on speech features (speech-BERT). We adopt the pre-training strategy of token level mask language modeling for speech-BERT as well, and further fine-tune with our zero-inflated layer for the mixture of discrete and continuous outputs. The experimental results show that our approach achieves better performance on WER prediction in the metrics of Pearson and MAE, compared with most existed quality estimation algorithms for ASR or machine translation.

Shiliang Zhang, Ming Lei, Zhijie Yan

Connectionist Temporal Classification (CTC) based end-to-end speech recognition system usually need to incorporate an external language model by using WFST-based decoding in order to achieve promising results. This is more essential to Mandarin speech recognition since it owns a special phenomenon, namely homophone, which causes a lot of substitution errors. The linguistic information introduced by language model will help to distinguish these substitution errors. In this work, we propose a transformer based spelling correction model to automatically correct errors especially the substitution errors made by CTC-based Mandarin speech recognition system. Specifically, we investigate using the recognition results generated by CTC-based systems as input and the ground-truth transcriptions as output to train a transformer with encoder-decoder architecture, which is much similar to machine translation. Results in a 20,000 hours Mandarin speech recognition task show that the proposed spelling correction model can achieve a CER of 3.41%, which results in 22.9% and 53.2% relative improvement compared to the baseline CTC-based systems decoded with and without language model respectively.

Zhiying Huang, Heng Lu, Ming Lei et al.

Speaker adaptation methods aim to create fair quality synthesis speech voice font for target speakers while only limited resources available. Recently, as deep neural networks based statistical parametric speech synthesis (SPSS) methods become dominant in SPSS TTS back-end modeling, speaker adaptation under the neural network based SPSS framework has also became an important task. In this paper, linear networks (LN) is inserted in multiple neural network layers and fine-tuned together with output layer for best speaker adaptation performance. When adaptation data is extremely small, the low-rank plus diagonal(LRPD) decomposition for LN is employed to make the adapted voice more stable. Speaker adaptation experiments are conducted under a range of adaptation utterances numbers. Moreover, speaker adaptation from 1) female to female, 2) male to female and 3) female to male are investigated. Objective measurement and subjective tests show that LN with LRPD decomposition performs most stable when adaptation data is extremely limited, and our best speaker adaptation (SA) model with only 200 adaptation utterances achieves comparable quality with speaker dependent (SD) model trained with 1000 utterances, in both naturalness and similarity to target speaker.

Shiliang Zhang, Ming Lei, Zhijie Yan et al.

In this paper, we present an improved feedforward sequential memory networks (FSMN) architecture, namely Deep-FSMN (DFSMN), by introducing skip connections between memory blocks in adjacent layers. These skip connections enable the information flow across different layers and thus alleviate the gradient vanishing problem when building very deep structure. As a result, DFSMN significantly benefits from these skip connections and deep structure. We have compared the performance of DFSMN to BLSTM both with and without lower frame rate (LFR) on several large speech recognition tasks, including English and Mandarin. Experimental results shown that DFSMN can consistently outperform BLSTM with dramatic gain, especially trained with LFR using CD-Phone as modeling units. In the 2000 hours Fisher (FSH) task, the proposed DFSMN can achieve a word error rate of 9.4% by purely using the cross-entropy criterion and decoding with a 3-gram language model, which achieves a 1.5% absolute improvement compared to the BLSTM. In a 20000 hours Mandarin recognition task, the LFR trained DFSMN can achieve more than 20% relative improvement compared to the LFR trained BLSTM. Moreover, we can easily design the lookahead filter order of the memory blocks in DFSMN to control the latency for real-time applications.

Mengxiao Bi, Heng Lu, Shiliang Zhang et al.

The Bidirectional LSTM (BLSTM) RNN based speech synthesis system is among the best parametric Text-to-Speech (TTS) systems in terms of the naturalness of generated speech, especially the naturalness in prosody. However, the model complexity and inference cost of BLSTM prevents its usage in many runtime applications. Meanwhile, Deep Feed-forward Sequential Memory Networks (DFSMN) has shown its consistent out-performance over BLSTM in both word error rate (WER) and the runtime computation cost in speech recognition tasks. Since speech synthesis also requires to model long-term dependencies compared to speech recognition, in this paper, we investigate the Deep-FSMN (DFSMN) in speech synthesis. Both objective and subjective experiments show that, compared with BLSTM TTS method, the DFSMN system can generate synthesized speech with comparable speech quality while drastically reduce model complexity and speech generation time.