Chao Weng

Haoxin Chen, Menghan Xia, Yingqing He et al.

Video generation has increasingly gained interest in both academia and industry. Although commercial tools can generate plausible videos, there is a limited number of open-source models available for researchers and engineers. In this work, we introduce two diffusion models for high-quality video generation, namely text-to-video (T2V) and image-to-video (I2V) models. T2V models synthesize a video based on a given text input, while I2V models incorporate an additional image input. Our proposed T2V model can generate realistic and cinematic-quality videos with a resolution of $1024 \times 576$, outperforming other open-source T2V models in terms of quality. The I2V model is designed to produce videos that strictly adhere to the content of the provided reference image, preserving its content, structure, and style. This model is the first open-source I2V foundation model capable of transforming a given image into a video clip while maintaining content preservation constraints. We believe that these open-source video generation models will contribute significantly to the technological advancements within the community.

Dongchao Yang, Jianwei Yu, Helin Wang et al.

Generating sound effects that humans want is an important topic. However, there are few studies in this area for sound generation. In this study, we investigate generating sound conditioned on a text prompt and propose a novel text-to-sound generation framework that consists of a text encoder, a Vector Quantized Variational Autoencoder (VQ-VAE), a decoder, and a vocoder. The framework first uses the decoder to transfer the text features extracted from the text encoder to a mel-spectrogram with the help of VQ-VAE, and then the vocoder is used to transform the generated mel-spectrogram into a waveform. We found that the decoder significantly influences the generation performance. Thus, we focus on designing a good decoder in this study. We begin with the traditional autoregressive decoder, which has been proved as a state-of-the-art method in previous sound generation works. However, the AR decoder always predicts the mel-spectrogram tokens one by one in order, which introduces the unidirectional bias and accumulation of errors problems. Moreover, with the AR decoder, the sound generation time increases linearly with the sound duration. To overcome the shortcomings introduced by AR decoders, we propose a non-autoregressive decoder based on the discrete diffusion model, named Diffsound. Specifically, the Diffsound predicts all of the mel-spectrogram tokens in one step and then refines the predicted tokens in the next step, so the best-predicted results can be obtained after several steps. Our experiments show that our proposed Diffsound not only produces better text-to-sound generation results when compared with the AR decoder but also has a faster generation speed, e.g., MOS: 3.56 \textit{v.s} 2.786, and the generation speed is five times faster than the AR decoder.

Yingqing He, Menghan Xia, Haoxin Chen et al. · tsinghua

Generating videos for visual storytelling can be a tedious and complex process that typically requires either live-action filming or graphics animation rendering. To bypass these challenges, our key idea is to utilize the abundance of existing video clips and synthesize a coherent storytelling video by customizing their appearances. We achieve this by developing a framework comprised of two functional modules: (i) Motion Structure Retrieval, which provides video candidates with desired scene or motion context described by query texts, and (ii) Structure-Guided Text-to-Video Synthesis, which generates plot-aligned videos under the guidance of motion structure and text prompts. For the first module, we leverage an off-the-shelf video retrieval system and extract video depths as motion structure. For the second module, we propose a controllable video generation model that offers flexible controls over structure and characters. The videos are synthesized by following the structural guidance and appearance instruction. To ensure visual consistency across clips, we propose an effective concept personalization approach, which allows the specification of the desired character identities through text prompts. Extensive experiments demonstrate that our approach exhibits significant advantages over various existing baselines.

Jinchuan Tian, Jianwei Yu, Chao Weng et al.

In automatic speech recognition (ASR) research, discriminative criteria have achieved superior performance in DNN-HMM systems. Given this success, the adoption of discriminative criteria is promising to boost the performance of end-to-end (E2E) ASR systems. With this motivation, previous works have introduced the minimum Bayesian risk (MBR, one of the discriminative criteria) into E2E ASR systems. However, the effectiveness and efficiency of the MBR-based methods are compromised: the MBR criterion is only used in system training, which creates a mismatch between training and decoding; the on-the-fly decoding process in MBR-based methods results in the need for pre-trained models and slow training speeds. To this end, novel algorithms are proposed in this work to integrate another widely used discriminative criterion, lattice-free maximum mutual information (LF-MMI), into E2E ASR systems not only in the training stage but also in the decoding process. The proposed LF-MMI training and decoding methods show their effectiveness on two widely used E2E frameworks: Attention-Based Encoder-Decoders (AEDs) and Neural Transducers (NTs). Compared with MBR-based methods, the proposed LF-MMI method: maintains the consistency between training and decoding; eschews the on-the-fly decoding process; trains from randomly initialized models with superior training efficiency. Experiments suggest that the LF-MMI method outperforms its MBR counterparts and consistently leads to statistically significant performance improvements on various frameworks and datasets from 30 hours to 14.3k hours. The proposed method achieves state-of-the-art (SOTA) results on Aishell-1 (CER 4.10%) and Aishell-2 (CER 5.02%) datasets. Code is released.

Jinchuan Tian, Jianwei Yu, Chunlei Zhang et al.

Despite the rapid progress in automatic speech recognition (ASR) research, recognizing multilingual speech using a unified ASR system remains highly challenging. Previous works on multilingual speech recognition mainly focus on two directions: recognizing multiple monolingual speech or recognizing code-switched speech that uses different languages interchangeably within a single utterance. However, a pragmatic multilingual recognizer is expected to be compatible with both directions. In this work, a novel language-aware encoder (LAE) architecture is proposed to handle both situations by disentangling language-specific information and generating frame-level language-aware representations during encoding. In the LAE, the primary encoding is implemented by the shared block while the language-specific blocks are used to extract specific representations for each language. To learn language-specific information discriminatively, a language-aware training method is proposed to optimize the language-specific blocks in LAE. Experiments conducted on Mandarin-English code-switched speech suggest that the proposed LAE is capable of discriminating different languages in frame-level and shows superior performance on both monolingual and multilingual ASR tasks. With either a real-recorded or simulated code-switched dataset, the proposed LAE achieves statistically significant improvements on both CTC and neural transducer systems. Code is released

Jinchuan Tian, Brian Yan, Jianwei Yu et al.

Sequence-to-Sequence (seq2seq) tasks transcribe the input sequence to a target sequence. The Connectionist Temporal Classification (CTC) criterion is widely used in multiple seq2seq tasks. Besides predicting the target sequence, a side product of CTC is to predict the alignment, which is the most probable input-long sequence that specifies a hard aligning relationship between the input and target units. As there are multiple potential aligning sequences (called paths) that are equally considered in CTC formulation, the choice of which path will be most probable and become the predicted alignment is always uncertain. In addition, it is usually observed that the alignment predicted by vanilla CTC will drift compared with its reference and rarely provides practical functionalities. Thus, the motivation of this work is to make the CTC alignment prediction controllable and thus equip CTC with extra functionalities. The Bayes risk CTC (BRCTC) criterion is then proposed in this work, in which a customizable Bayes risk function is adopted to enforce the desired characteristics of the predicted alignment. With the risk function, the BRCTC is a general framework to adopt some customizable preference over the paths in order to concentrate the posterior into a particular subset of the paths. In applications, we explore one particular preference which yields models with the down-sampling ability and reduced inference costs. By using BRCTC with another preference for early emissions, we obtain an improved performance-latency trade-off for online models. Experimentally, the proposed BRCTC reduces the inference cost of offline models by up to 47% without performance degradation and cuts down the overall latency of online systems to an unseen level.

Hangting Chen, Jianwei Yu, Yi Luo et al.

Echo cancellation and noise reduction are essential for full-duplex communication, yet most existing neural networks have high computational costs and are inflexible in tuning model complexity. In this paper, we introduce time-frequency dual-path compression to achieve a wide range of compression ratios on computational cost. Specifically, for frequency compression, trainable filters are used to replace manually designed filters for dimension reduction. For time compression, only using frame skipped prediction causes large performance degradation, which can be alleviated by a post-processing network with full sequence modeling. We have found that under fixed compression ratios, dual-path compression combining both the time and frequency methods will give further performance improvement, covering compression ratios from 4x to 32x with little model size change. Moreover, the proposed models show competitive performance compared with fast FullSubNet and DeepFilterNet.

Jinchuan Tian, Jianwei Yu, Hangting Chen et al.

Automatic speech recognition (ASR) based on transducers is widely used. In training, a transducer maximizes the summed posteriors of all paths. The path with the highest posterior is commonly defined as the predicted alignment between the speech and the transcription. While the vanilla transducer does not have a prior preference for any of the valid paths, this work intends to enforce the preferred paths and achieve controllable alignment prediction. Specifically, this work proposes Bayes Risk Transducer (BRT), which uses a Bayes risk function to set lower risk values to the preferred paths so that the predicted alignment is more likely to satisfy specific desired properties. We further demonstrate that these predicted alignments with intentionally designed properties can provide practical advantages over the vanilla transducer. Experimentally, the proposed BRT saves inference cost by up to 46% for non-streaming ASR and reduces overall system latency by 41% for streaming ASR.

Boyi Kang, Xinfa Zhu, Zihan Zhang et al.

Recent advancements in language models (LMs) have demonstrated strong capabilities in semantic understanding and contextual modeling, which have flourished in generative speech enhancement (SE). However, many LM-based SE approaches primarily focus on semantic information, often neglecting the critical role of acoustic information, which leads to acoustic inconsistency after enhancement and limited generalization across diverse SE tasks. In this paper, we introduce LLaSE-G1, a LLaMA-based language model that incentivizes generalization capabilities for speech enhancement. LLaSE-G1 offers the following key contributions: First, to mitigate acoustic inconsistency, LLaSE-G1 employs continuous representations from WavLM as input and predicts speech tokens from X-Codec2, maximizing acoustic preservation. Second, to promote generalization capability, LLaSE-G1 introduces dual-channel inputs and outputs, unifying multiple SE tasks without requiring task-specific IDs. Third, LLaSE-G1 outperforms prior task-specific discriminative and generative SE models, demonstrating scaling effects at test time and emerging capabilities for unseen SE tasks. Additionally, we release our code and models to support further research in this area.

Xin He, Shaoli Huang, Xiaohang Zhan et al.

Current techniques face difficulties in generating motions from intricate semantic descriptions, primarily due to insufficient semantic annotations in datasets and weak contextual understanding. To address these issues, we present SemanticBoost, a novel framework that tackles both challenges simultaneously. Our framework comprises a Semantic Enhancement module and a Context-Attuned Motion Denoiser (CAMD). The Semantic Enhancement module extracts supplementary semantics from motion data, enriching the dataset's textual description and ensuring precise alignment between text and motion data without depending on large language models. On the other hand, the CAMD approach provides an all-encompassing solution for generating high-quality, semantically consistent motion sequences by effectively capturing context information and aligning the generated motion with the given textual descriptions. Distinct from existing methods, our approach can synthesize accurate orientational movements, combined motions based on specific body part descriptions, and motions generated from complex, extended sentences. Our experimental results demonstrate that SemanticBoost, as a diffusion-based method, outperforms auto-regressive-based techniques, achieving cutting-edge performance on the Humanml3D dataset while maintaining realistic and smooth motion generation quality.

Haoxin Chen, Yong Zhang, Xiaodong Cun et al.

Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition.

Hao Yan, Xingchen Liu, Hao Wang et al.

Recent strides in multimodal large language models (MLLMs) have significantly advanced their performance in many reasoning tasks. However, Abstract Visual Reasoning (AVR) remains a critical challenge, primarily due to limitations in perceiving abstract graphics. To tackle this issue, we investigate the bottlenecks in current MLLMs and synthesize training data to improve their abstract visual perception. First, we propose VisuRiddles, a benchmark for AVR, featuring tasks meticulously constructed to assess models' reasoning capacities across five core dimensions and two high-level reasoning categories. Second, we introduce the Perceptual Riddle Synthesizer (PRS), an automated framework for generating riddles with fine-grained perceptual descriptions. PRS not only generates valuable training data for abstract graphics but also provides fine-grained perceptual description, crucially allowing for supervision over intermediate reasoning stages and thereby improving both training efficacy and model interpretability. Our extensive experimental results on VisuRiddles empirically validate that fine-grained visual perception is the principal bottleneck and our synthesis framework markedly enhances the performance of contemporary MLLMs on these challenging tasks. Our code and dataset will be released at https://github.com/yh-hust/VisuRiddles

Jiatong Shi, Jionghao Han, Yichen Lu et al.

Role-play has become a key testbed for generative models, expanding from text-only dialogue to multimodal interaction. Extending role-play to speech captures prosody, emotion, and delivery, but also poses new evaluation challenges. Current pipelines often use audio large language models (ALLMs) as zero-shot judges, which miss paralinguistic cues, collapse multiple aspects into coarse scores, and rely on synthetic speech references that fail to reflect real-world roles. We present Speech-DRAME, a unified framework that contributes at three levels: (i) Speech-DRAME-EvalBench, an evaluation benchmark with bilingual human-annotated data and protocols for training and testing speech evaluation models (SEMs), (ii) DRAME-Eval, a fine-tuned evaluation model, which substantially outperforms zero-shot and few-shot ALLMs, and (iii) Speech-DRAME-RoleBench, a speech role-play benchmark that leverages DRAME-Eval as an automatic judge to compare speech foundation models (SFMs). Speech-DRAME distinguishes between two complementary evaluation strategies: Archetype Evaluation, a top-down approach measuring adherence to broad role archetypes, and Realism Evaluation, a bottom-up approach grounded in real human speech that emphasizes nuanced role quality. Compared to zero-shot ALLM judges, DRAME-Eval achieves stronger agreement with human ratings (Pearson correlation from 0.480 to 0.629 in archetypes, and 0.390 to 0.625 in realism). By integrating transparent benchmark resources, modeling approaches, and system-level evaluation, Speech-DRAME provides the first comprehensive, reproducible foundation for assessing spoken role-play.

Yi Luo, Jianwei Yu, Hangting Chen et al.

We introduce Gull, a generative multifunctional audio codec. Gull is a general purpose neural audio compression and decompression model which can be applied to a wide range of tasks and applications such as real-time communication, audio super-resolution, and codec language models. The key components of Gull include (1) universal-sample-rate modeling via subband modeling schemes motivated by recent progress in audio source separation, (2) gain-shape representations motivated by traditional audio codecs, (3) improved residual vector quantization modules, (4) elastic decoder network that enables user-defined model size and complexity during inference time, (5) built-in ability for audio super-resolution without the increase of bitrate. We compare Gull with existing traditional and neural audio codecs and show that Gull is able to achieve on par or better performance across various sample rates, bitrates and model complexities in both subjective and objective evaluation metrics.

Jianwei Cui, Yu Gu, Chao Weng et al.

This paper presents an advanced end-to-end singing voice synthesis (SVS) system based on the source-filter mechanism that directly translates lyrical and melodic cues into expressive and high-fidelity human-like singing. Similarly to VISinger 2, the proposed system also utilizes training paradigms evolved from VITS and incorporates elements like the fundamental pitch (F0) predictor and waveform generation decoder. To address the issue that the coupling of mel-spectrogram features with F0 information may introduce errors during F0 prediction, we consider two strategies. Firstly, we leverage mel-cepstrum (mcep) features to decouple the intertwined mel-spectrogram and F0 characteristics. Secondly, inspired by the neural source-filter models, we introduce source excitation signals as the representation of F0 in the SVS system, aiming to capture pitch nuances more accurately. Meanwhile, differentiable mcep and F0 losses are employed as the waveform decoder supervision to fortify the prediction accuracy of speech envelope and pitch in the generated speech. Experiments on the Opencpop dataset demonstrate efficacy of the proposed model in synthesis quality and intonation accuracy.

Jun Chen, Shichao Hu, Jiuxin Lin et al.

In-car multi-zone speech separation, which captures voices from different speech zones, plays a crucial role in human-vehicle interaction. Although previous SpatialNet has achieved notable results, its high computational cost still hinders real-time applications in vehicles. To this end, this paper proposes LSZone, a lightweight spatial information modeling architecture for real-time in-car multi-zone speech separation. We design a spatial information extraction-compression (SpaIEC) module that combines Mel spectrogram and Interaural Phase Difference (IPD) to reduce computational burden while maintaining performance. Additionally, to efficiently model spatial information, we introduce an extremely lightweight Conv-GRU crossband-narrowband processing (CNP) module. Experimental results demonstrate that LSZone, with a complexity of 0.56G MACs and a real-time factor (RTF) of 0.37, delivers impressive performance in complex noise and multi-speaker scenarios.

Rongjie Huang, Chunlei Zhang, Yongqi Wang et al.

Various applications of voice synthesis have been developed independently despite the fact that they generate "voice" as output in common. In addition, the majority of voice synthesis models currently rely on annotated audio data, but it is crucial to scale them to self-supervised datasets in order to effectively capture the wide range of acoustic variations present in human voice, including speaker identity, emotion, and prosody. In this work, we propose Make-A-Voice, a unified framework for synthesizing and manipulating voice signals from discrete representations. Make-A-Voice leverages a "coarse-to-fine" approach to model the human voice, which involves three stages: 1) semantic stage: model high-level transformation between linguistic content and self-supervised semantic tokens, 2) acoustic stage: introduce varying control signals as acoustic conditions for semantic-to-acoustic modeling, and 3) generation stage: synthesize high-fidelity waveforms from acoustic tokens. Make-A-Voice offers notable benefits as a unified voice synthesis framework: 1) Data scalability: the major backbone (i.e., acoustic and generation stage) does not require any annotations, and thus the training data could be scaled up. 2) Controllability and conditioning flexibility: we investigate different conditioning mechanisms and effectively handle three voice synthesis applications, including text-to-speech (TTS), voice conversion (VC), and singing voice synthesis (SVS) by re-synthesizing the discrete voice representations with prompt guidance. Experimental results demonstrate that Make-A-Voice exhibits superior audio quality and style similarity compared with competitive baseline models. Audio samples are available at https://Make-A-Voice.github.io

Naijun Zheng, Na Li, Xixin Wu et al.

This paper describes our speaker diarization system submitted to the Multi-channel Multi-party Meeting Transcription (M2MeT) challenge, where Mandarin meeting data were recorded in multi-channel format for diarization and automatic speech recognition (ASR) tasks. In these meeting scenarios, the uncertainty of the speaker number and the high ratio of overlapped speech present great challenges for diarization. Based on the assumption that there is valuable complementary information between acoustic features, spatial-related and speaker-related features, we propose a multi-level feature fusion mechanism based target-speaker voice activity detection (FFM-TS-VAD) system to improve the performance of the conventional TS-VAD system. Furthermore, we propose a data augmentation method during training to improve the system robustness when the angular difference between two speakers is relatively small. We provide comparisons for different sub-systems we used in M2MeT challenge. Our submission is a fusion of several sub-systems and ranks second in the diarization task.

Jinchuan Tian, Jianwei Yu, Chao Weng et al.

Despite the rapid progress of end-to-end (E2E) automatic speech recognition (ASR), it has been shown that incorporating external language models (LMs) into the decoding can further improve the recognition performance of E2E ASR systems. To align with the modeling units adopted in E2E ASR systems, subword-level (e.g., characters, BPE) LMs are usually used to cooperate with current E2E ASR systems. However, the use of subword-level LMs will ignore the word-level information, which may limit the strength of the external LMs in E2E ASR. Although several methods have been proposed to incorporate word-level external LMs in E2E ASR, these methods are mainly designed for languages with clear word boundaries such as English and cannot be directly applied to languages like Mandarin, in which each character sequence can have multiple corresponding word sequences. To this end, we propose a novel decoding algorithm where a word-level lattice is constructed on-the-fly to consider all possible word sequences for each partial hypothesis. Then, the LM score of the hypothesis is obtained by intersecting the generated lattice with an external word N-gram LM. The proposed method is examined on both Attention-based Encoder-Decoder (AED) and Neural Transducer (NT) frameworks. Experiments suggest that our method consistently outperforms subword-level LMs, including N-gram LM and neural network LM. We achieve state-of-the-art results on both Aishell-1 (CER 4.18%) and Aishell-2 (CER 5.06%) datasets and reduce CER by 14.8% relatively on a 21K-hour Mandarin dataset.

Jinchuan Tian, Jianwei Yu, Chao Weng et al.

Recently, End-to-End (E2E) frameworks have achieved remarkable results on various Automatic Speech Recognition (ASR) tasks. However, Lattice-Free Maximum Mutual Information (LF-MMI), as one of the discriminative training criteria that show superior performance in hybrid ASR systems, is rarely adopted in E2E ASR frameworks. In this work, we propose a novel approach to integrate LF-MMI criterion into E2E ASR frameworks in both training and decoding stages. The proposed approach shows its effectiveness on two of the most widely used E2E frameworks including Attention-Based Encoder-Decoders (AEDs) and Neural Transducers (NTs). Experiments suggest that the introduction of the LF-MMI criterion consistently leads to significant performance improvements on various datasets and different E2E ASR frameworks. The best of our models achieves competitive CER of 4.1\% / 4.4\% on Aishell-1 dev/test set; we also achieve significant error reduction on Aishell-2 and Librispeech datasets over strong baselines.

Brian Yan, Chunlei Zhang, Meng Yu et al.

Conversational bilingual speech encompasses three types of utterances: two purely monolingual types and one intra-sententially code-switched type. In this work, we propose a general framework to jointly model the likelihoods of the monolingual and code-switch sub-tasks that comprise bilingual speech recognition. By defining the monolingual sub-tasks with label-to-frame synchronization, our joint modeling framework can be conditionally factorized such that the final bilingual output, which may or may not be code-switched, is obtained given only monolingual information. We show that this conditionally factorized joint framework can be modeled by an end-to-end differentiable neural network. We demonstrate the efficacy of our proposed model on bilingual Mandarin-English speech recognition across both monolingual and code-switched corpora.

Xiaoyi Qin, Na Li, Chao Weng et al.

Recently, the attention mechanism such as squeeze-and-excitation module (SE) and convolutional block attention module (CBAM) has achieved great success in deep learning-based speaker verification system. This paper introduces an alternative effective yet simple one, i.e., simple attention module (SimAM), for speaker verification. The SimAM module is a plug-and-play module without extra modal parameters. In addition, we propose a noisy label detection method to iteratively filter out the data samples with a noisy label from the training data, considering that a large-scale dataset labeled with human annotation or other automated processes may contain noisy labels. Data with the noisy label may over parameterize a deep neural network (DNN) and result in a performance drop due to the memorization effect of the DNN. Experiments are conducted on VoxCeleb dataset. The speaker verification model with SimAM achieves the 0.675% equal error rate (EER) on VoxCeleb1 original test trials. Our proposed iterative noisy label detection method further reduces the EER to 0.643%.

Guoguo Chen, Shuzhou Chai, Guanbo Wang et al.

This paper introduces GigaSpeech, an evolving, multi-domain English speech recognition corpus with 10,000 hours of high quality labeled audio suitable for supervised training, and 40,000 hours of total audio suitable for semi-supervised and unsupervised training. Around 40,000 hours of transcribed audio is first collected from audiobooks, podcasts and YouTube, covering both read and spontaneous speaking styles, and a variety of topics, such as arts, science, sports, etc. A new forced alignment and segmentation pipeline is proposed to create sentence segments suitable for speech recognition training, and to filter out segments with low-quality transcription. For system training, GigaSpeech provides five subsets of different sizes, 10h, 250h, 1000h, 2500h, and 10000h. For our 10,000-hour XL training subset, we cap the word error rate at 4% during the filtering/validation stage, and for all our other smaller training subsets, we cap it at 0%. The DEV and TEST evaluation sets, on the other hand, are re-processed by professional human transcribers to ensure high transcription quality. Baseline systems are provided for popular speech recognition toolkits, namely Athena, ESPnet, Kaldi and Pika.

Jingbei Li, Yi Meng, Chenyi Li et al.

Comparing with traditional text-to-speech (TTS) systems, conversational TTS systems are required to synthesize speeches with proper speaking style confirming to the conversational context. However, state-of-the-art context modeling methods in conversational TTS only model the textual information in context with a recurrent neural network (RNN). Such methods have limited ability in modeling the inter-speaker influence in conversations, and also neglect the speaking styles and the intra-speaker inertia inside each speaker. Inspired by DialogueGCN and its superiority in modeling such conversational influences than RNN based approaches, we propose a graph-based multi-modal context modeling method and adopt it to conversational TTS to enhance the speaking styles of synthesized speeches. Both the textual and speaking style information in the context are extracted and processed by DialogueGCN to model the inter- and intra-speaker influence in conversations. The outputs of DialogueGCN are then summarized by attention mechanism, and converted to the enhanced speaking style for current utterance. An English conversation corpus is collected and annotated for our research and released to public. Experiment results on this corpus demonstrate the effectiveness of our proposed approach, which outperforms the state-of-the-art context modeling method in conversational TTS in both MOS and ABX preference rate.

Max W. Y. Lam, Jun Wang, Chao Weng et al.

End-to-end speech recognition generally uses hand-engineered acoustic features as input and excludes the feature extraction module from its joint optimization. To extract learnable and adaptive features and mitigate information loss, we propose a new encoder that adopts globally attentive locally recurrent (GALR) networks and directly takes raw waveform as input. We observe improved ASR performance and robustness by applying GALR on different window lengths to aggregate fine-grain temporal information into multi-scale acoustic features. Experiments are conducted on a benchmark dataset AISHELL-2 and two large-scale Mandarin speech corpus of 5,000 hours and 21,000 hours. With faster speed and comparable model size, our proposed multi-scale GALR waveform encoder achieved consistent character error rate reductions (CERRs) from 7.9% to 28.1% relative over strong baselines, including Conformer and TDNN-Conformer. In particular, our approach demonstrated notable robustness than the traditional handcrafted features and outperformed the baseline MFCC-based TDNN-Conformer model by a 15.2% CERR on a music-mixed real-world speech test set.

Helin Wang, Bo Wu, Lianwu Chen et al.

In this paper, we exploit the effective way to leverage contextual information to improve the speech dereverberation performance in real-world reverberant environments. We propose a temporal-contextual attention approach on the deep neural network (DNN) for environment-aware speech dereverberation, which can adaptively attend to the contextual information. More specifically, a FullBand based Temporal Attention approach (FTA) is proposed, which models the correlations between the fullband information of the context frames. In addition, considering the difference between the attenuation of high frequency bands and low frequency bands (high frequency bands attenuate faster than low frequency bands) in the room impulse response (RIR), we also propose a SubBand based Temporal Attention approach (STA). In order to guide the network to be more aware of the reverberant environments, we jointly optimize the dereverberation network and the reverberation time (RT60) estimator in a multi-task manner. Our experimental results indicate that the proposed method outperforms our previously proposed reverberation-time-aware DNN and the learned attention weights are fully physical consistent. We also report a preliminary yet promising dereverberation and recognition experiment on real test data.

Aswin Shanmugam Subramanian, Chao Weng, Shinji Watanabe et al.

Multi-source localization is an important and challenging technique for multi-talker conversation analysis. This paper proposes a novel supervised learning method using deep neural networks to estimate the direction of arrival (DOA) of all the speakers simultaneously from the audio mixture. At the heart of the proposal is a source splitting mechanism that creates source-specific intermediate representations inside the network. This allows our model to give source-specific posteriors as the output unlike the traditional multi-label classification approach. Existing deep learning methods perform a frame level prediction, whereas our approach performs an utterance level prediction by incorporating temporal selection and averaging inside the network to avoid post-processing. We also experiment with various loss functions and show that a variant of earth mover distance (EMD) is very effective in classifying DOA at a very high resolution by modeling inter-class relationships. In addition to using the prediction error as a metric for evaluating our localization model, we also establish its potency as a frontend with automatic speech recognition (ASR) as the downstream task. We convert the estimated DOAs into a feature suitable for ASR and pass it as an additional input feature to a strong multi-channel and multi-talker speech recognition baseline. This added input feature drastically improves the ASR performance and gives a word error rate (WER) of 6.3% on the evaluation data of our simulated noisy two speaker mixtures, while the baseline which doesn't use explicit localization input has a WER of 11.5%. We also perform ASR evaluation on real recordings with the overlapped set of the MC-WSJ-AV corpus in addition to simulated mixtures.

Peng Liu, Yuewen Cao, Songxiang Liu et al.

This paper proposes VARA-TTS, a non-autoregressive (non-AR) text-to-speech (TTS) model using a very deep Variational Autoencoder (VDVAE) with Residual Attention mechanism, which refines the textual-to-acoustic alignment layer-wisely. Hierarchical latent variables with different temporal resolutions from the VDVAE are used as queries for residual attention module. By leveraging the coarse global alignment from previous attention layer as an extra input, the following attention layer can produce a refined version of alignment. This amortizes the burden of learning the textual-to-acoustic alignment among multiple attention layers and outperforms the use of only a single attention layer in robustness. An utterance-level speaking speed factor is computed by a jointly-trained speaking speed predictor, which takes the mean-pooled latent variables of the coarsest layer as input, to determine number of acoustic frames at inference. Experimental results show that VARA-TTS achieves slightly inferior speech quality to an AR counterpart Tacotron 2 but an order-of-magnitude speed-up at inference; and outperforms an analogous non-AR model, BVAE-TTS, in terms of speech quality.

Wei Xia, Chunlei Zhang, Chao Weng et al.

In this study, we investigate self-supervised representation learning for speaker verification (SV). First, we examine a simple contrastive learning approach (SimCLR) with a momentum contrastive (MoCo) learning framework, where the MoCo speaker embedding system utilizes a queue to maintain a large set of negative examples. We show that better speaker embeddings can be learned by momentum contrastive learning. Next, alternative augmentation strategies are explored to normalize extrinsic speaker variabilities of two random segments from the same speech utterance. Specifically, augmentation in the waveform largely improves the speaker representations for SV tasks. The proposed MoCo speaker embedding is further improved when a prototypical memory bank is introduced, which encourages the speaker embeddings to be closer to their assigned prototypes with an intermediate clustering step. In addition, we generalize the self-supervised framework to a semi-supervised scenario where only a small portion of the data is labeled. Comprehensive experiments on the Voxceleb dataset demonstrate that our proposed self-supervised approach achieves competitive performance compared with existing techniques, and can approach fully supervised results with partially labeled data.

Jiatong Shi, Chunlei Zhang, Chao Weng et al.

Target-speaker speech recognition aims to recognize target-speaker speech from noisy environments with background noise and interfering speakers. This work presents a joint framework that combines time-domain target-speaker speech extraction and Recurrent Neural Network Transducer (RNN-T). To stabilize the joint-training, we propose a multi-stage training strategy that pre-trains and fine-tunes each module in the system before joint-training. Meanwhile, speaker identity and speech enhancement uncertainty measures are proposed to compensate for residual noise and artifacts from the target speech extraction module. Compared to a recognizer fine-tuned with a target speech extraction model, our experiments show that adding the neural uncertainty module significantly reduces 17% relative Character Error Rate (CER) on multi-speaker signals with background noise. The multi-condition experiments indicate that our method can achieve 9% relative performance gain in the noisy condition while maintaining the performance in the clean condition.

Aswin Shanmugam Subramanian, Chao Weng, Shinji Watanabe et al.

This paper proposes a new paradigm for handling far-field multi-speaker data in an end-to-end neural network manner, called directional automatic speech recognition (D-ASR), which explicitly models source speaker locations. In D-ASR, the azimuth angle of the sources with respect to the microphone array is defined as a latent variable. This angle controls the quality of separation, which in turn determines the ASR performance. All three functionalities of D-ASR: localization, separation, and recognition are connected as a single differentiable neural network and trained solely based on ASR error minimization objectives. The advantages of D-ASR over existing methods are threefold: (1) it provides explicit speaker locations, (2) it improves the explainability factor, and (3) it achieves better ASR performance as the process is more streamlined. In addition, D-ASR does not require explicit direction of arrival (DOA) supervision like existing data-driven localization models, which makes it more appropriate for realistic data. For the case of two source mixtures, D-ASR achieves an average DOA prediction error of less than three degrees. It also outperforms a strong far-field multi-speaker end-to-end system in both separation quality and ASR performance.

Xingchen Song, Zhiyong Wu, Yiheng Huang et al.

Non-autoregressive (NAR) transformer models have achieved significantly inference speedup but at the cost of inferior accuracy compared to autoregressive (AR) models in automatic speech recognition (ASR). Most of the NAR transformers take a fixed-length sequence filled with MASK tokens or a redundant sequence copied from encoder states as decoder input, they cannot provide efficient target-side information thus leading to accuracy degradation. To address this problem, we propose a CTC-enhanced NAR transformer, which generates target sequence by refining predictions of the CTC module. Experimental results show that our method outperforms all previous NAR counterparts and achieves 50x faster decoding speed than a strong AR baseline with only 0.0 ~ 0.3 absolute CER degradation on Aishell-1 and Aishell-2 datasets.

Xu Li, Na Li, Chao Weng et al.

Existing approaches for replay and synthetic speech detection still lack generalizability to unseen spoofing attacks. This work proposes to leverage a novel model structure, so-called Res2Net, to improve the anti-spoofing countermeasure's generalizability. Res2Net mainly modifies the ResNet block to enable multiple feature scales. Specifically, it splits the feature maps within one block into multiple channel groups and designs a residual-like connection across different channel groups. Such connection increases the possible receptive fields, resulting in multiple feature scales. This multiple scaling mechanism significantly improves the countermeasure's generalizability to unseen spoofing attacks. It also decreases the model size compared to ResNet-based models. Experimental results show that the Res2Net model consistently outperforms ResNet34 and ResNet50 by a large margin in both physical access (PA) and logical access (LA) of the ASVspoof 2019 corpus. Moreover, integration with the squeeze-and-excitation (SE) block can further enhance performance. For feature engineering, we investigate the generalizability of Res2Net combined with different acoustic features, and observe that the constant-Q transform (CQT) achieves the most promising performance in both PA and LA scenarios. Our best single system outperforms other state-of-the-art single systems in both PA and LA of the ASVspoof 2019 corpus.

Yusong Wu, Shengchen Li, Chengzhu Yu et al.

Peking Opera has been the most dominant form of Chinese performing art since around 200 years ago. A Peking Opera singer usually exhibits a very strong personal style via introducing improvisation and expressiveness on stage which leads the actual rhythm and pitch contour to deviate significantly from the original music score. This inconsistency poses a great challenge in Peking Opera singing voice synthesis from a music score. In this work, we propose to deal with this issue and synthesize expressive Peking Opera singing from the music score based on the Duration Informed Attention Network (DurIAN) framework. To tackle the rhythm mismatch, Lagrange multiplier is used to find the optimal output phoneme duration sequence with the constraint of the given note duration from music score. As for the pitch contour mismatch, instead of directly inferring from music score, we adopt a pseudo music score generated from the real singing and feed it as input during training. The experiments demonstrate that with the proposed system we can synthesize Peking Opera singing voice with high-quality timbre, pitch and expressiveness.

Liqiang Zhang, Chengzhu Yu, Heng Lu et al.

Singing voice conversion is converting the timbre in the source singing to the target speaker's voice while keeping singing content the same. However, singing data for target speaker is much more difficult to collect compared with normal speech data.In this paper, we introduce a singing voice conversion algorithm that is capable of generating high quality target speaker's singing using only his/her normal speech data. First, we manage to integrate the training and conversion process of speech and singing into one framework by unifying the features used in standard speech synthesis system and singing synthesis system. In this way, normal speech data can also contribute to singing voice conversion training, making the singing voice conversion system more robust especially when the singing database is small.Moreover, in order to achieve one-shot singing voice conversion, a speaker embedding module is developed using both speech and singing data, which provides target speaker identify information during conversion. Experiments indicate proposed sing conversion system can convert source singing to target speaker's high-quality singing with only 20 seconds of target speaker's enrollment speech data.

Yong Xu, Meng Yu, Shi-Xiong Zhang et al.

Purely neural network (NN) based speech separation and enhancement methods, although can achieve good objective scores, inevitably cause nonlinear speech distortions that are harmful for the automatic speech recognition (ASR). On the other hand, the minimum variance distortionless response (MVDR) beamformer with NN-predicted masks, although can significantly reduce speech distortions, has limited noise reduction capability. In this paper, we propose a multi-tap MVDR beamformer with complex-valued masks for speech separation and enhancement. Compared to the state-of-the-art NN-mask based MVDR beamformer, the multi-tap MVDR beamformer exploits the inter-frame correlation in addition to the inter-microphone correlation that is already utilized in prior arts. Further improvements include the replacement of the real-valued masks with the complex-valued masks and the joint training of the complex-mask NN. The evaluation on our multi-modal multi-channel target speech separation and enhancement platform demonstrates that our proposed multi-tap MVDR beamformer improves both the ASR accuracy and the perceptual speech quality against prior arts.

Yusong Wu, Shengchen Li, Chengzhu Yu et al.

This paper presents a method that generates expressive singing voice of Peking opera. The synthesis of expressive opera singing usually requires pitch contours to be extracted as the training data, which relies on techniques and is not able to be manually labeled. With the Duration Informed Attention Network (DurIAN), this paper makes use of musical note instead of pitch contours for expressive opera singing synthesis. The proposed method enables human annotation being combined with automatic extracted features to be used as training data thus the proposed method gives extra flexibility in data collection for Peking opera singing synthesis. Comparing with the expressive singing voice of Peking opera synthesised by pitch contour based system, the proposed musical note based system produces comparable singing voice in Peking opera with expressiveness in various aspects.

Liqiang Zhang, Chengzhu Yu, Heng Lu et al.

We propose an algorithm that is capable of synthesizing high quality target speaker's singing voice given only their normal speech samples. The proposed algorithm first integrate speech and singing synthesis into a unified framework, and learns universal speaker embeddings that are shareable between speech and singing synthesis tasks. Specifically, the speaker embeddings learned from normal speech via the speech synthesis objective are shared with those learned from singing samples via the singing synthesis objective in the unified training framework. This makes the learned speaker embedding a transferable representation for both speaking and singing. We evaluate the proposed algorithm on singing voice conversion task where the content of original singing is covered with the timbre of another speaker's voice learned purely from their normal speech samples. Our experiments indicate that the proposed algorithm generates high-quality singing voices that sound highly similar to target speaker's voice given only his or her normal speech samples. We believe that proposed algorithm will open up new opportunities for singing synthesis and conversion for broader users and applications.

Chengqi Deng, Chengzhu Yu, Heng Lu et al.

Singing voice conversion is to convert a singer's voice to another one's voice without changing singing content. Recent work shows that unsupervised singing voice conversion can be achieved with an autoencoder-based approach [1]. However, the converted singing voice can be easily out of key, showing that the existing approach cannot model the pitch information precisely. In this paper, we propose to advance the existing unsupervised singing voice conversion method proposed in [1] to achieve more accurate pitch translation and flexible pitch manipulation. Specifically, the proposed PitchNet added an adversarially trained pitch regression network to enforce the encoder network to learn pitch invariant phoneme representation, and a separate module to feed pitch extracted from the source audio to the decoder network. Our evaluation shows that the proposed method can greatly improve the quality of the converted singing voice (2.92 vs 3.75 in MOS). We also demonstrate that the pitch of converted singing can be easily controlled during generation by changing the levels of the extracted pitch before passing it to the decoder network.

Chao Weng, Chengzhu Yu, Jia Cui et al.

In this work, we propose minimum Bayes risk (MBR) training of RNN-Transducer (RNN-T) for end-to-end speech recognition. Specifically, initialized with a RNN-T trained model, MBR training is conducted via minimizing the expected edit distance between the reference label sequence and on-the-fly generated N-best hypothesis. We also introduce a heuristic to incorporate an external neural network language model (NNLM) in RNN-T beam search decoding and explore MBR training with the external NNLM. Experimental results demonstrate an MBR trained model outperforms a RNN-T trained model substantially and further improvements can be achieved if trained with an external NNLM. Our best MBR trained system achieves absolute character error rate (CER) reductions of 1.2% and 0.5% on read and spontaneous Mandarin speech respectively over a strong convolution and transformer based RNN-T baseline trained on ~21,000 hours of speech.

Zhao You, Dan Su, Jie Chen et al.

Self-attention networks (SAN) have been introduced into automatic speech recognition (ASR) and achieved state-of-the-art performance owing to its superior ability in capturing long term dependency. One of the key ingredients is the self-attention mechanism which can be effectively performed on the whole utterance level. In this paper, we try to investigate whether even more information beyond the whole utterance level can be exploited and beneficial. We propose to apply self-attention layer with augmented memory to ASR. Specifically, we first propose a variant model architecture which combines deep feed-forward sequential memory network (DFSMN) with self-attention layers to form a better baseline model compared with a purely self-attention network. Then, we propose and compare two kinds of additional memory structures added into self-attention layers. Experiments on large-scale LVCSR tasks show that on four individual test sets, the DFSMN-SAN architecture outperforms vanilla SAN encoder by 5% relatively in character error rate (CER). More importantly, the additional memory structure provides further 5% to 11% relative improvement in CER.

Chengzhu Yu, Heng Lu, Na Hu et al.

In this paper, we present a generic and robust multimodal synthesis system that produces highly natural speech and facial expression simultaneously. The key component of this system is the Duration Informed Attention Network (DurIAN), an autoregressive model in which the alignments between the input text and the output acoustic features are inferred from a duration model. This is different from the end-to-end attention mechanism used, and accounts for various unavoidable artifacts, in existing end-to-end speech synthesis systems such as Tacotron. Furthermore, DurIAN can be used to generate high quality facial expression which can be synchronized with generated speech with/without parallel speech and face data. To improve the efficiency of speech generation, we also propose a multi-band parallel generation strategy on top of the WaveRNN model. The proposed Multi-band WaveRNN effectively reduces the total computational complexity from 9.8 to 5.5 GFLOPS, and is able to generate audio that is 6 times faster than real time on a single CPU core. We show that DurIAN could generate highly natural speech that is on par with current state of the art end-to-end systems, while at the same time avoid word skipping/repeating errors in those systems. Finally, a simple yet effective approach for fine-grained control of expressiveness of speech and facial expression is introduced.

Chao Weng, Dong Yu

In this work, three lattice-free (LF) discriminative training criteria for purely sequence-trained neural network acoustic models are compared on LVCSR tasks, namely maximum mutual information (MMI), boosted maximum mutual information (bMMI) and state-level minimum Bayes risk (sMBR). We demonstrate that, analogous to LF-MMI, a neural network acoustic model can also be trained from scratch using LF-bMMI or LF-sMBR criteria respectively without the need of cross-entropy pre-training. Furthermore, experimental results on Switchboard-300hrs and Switchboard+Fisher-2100hrs datasets show that models trained with LF-bMMI consistently outperform those trained with plain LF-MMI and achieve a relative word error rate (WER) reduction of 5% over competitive temporal convolution projected LSTM (TDNN-LSTMP) LF-MMI baselines.