Zengwei Yao

Wei Kang, Liyong Guo, Fangjun Kuang et al. · nvidia

The transducer architecture is becoming increasingly popular in the field of speech recognition, because it is naturally streaming as well as high in accuracy. One of the drawbacks of transducer is that it is difficult to decode in a fast and parallel way due to an unconstrained number of symbols that can be emitted per time step. In this work, we introduce a constrained version of transducer loss to learn strictly monotonic alignments between the sequences; we also improve the standard greedy search and beam search algorithms by limiting the number of symbols that can be emitted per time step in transducer decoding, making it more efficient to decode in parallel with batches. Furthermore, we propose an finite state automaton-based (FSA) parallel beam search algorithm that can run with graphs on GPU efficiently. The experiment results show that we achieve slight word error rate (WER) improvement as well as significant speedup in decoding. Our work is open-sourced and publicly available\footnote{https://github.com/k2-fsa/icefall}.

Wei Kang, Zengwei Yao, Fangjun Kuang et al. · nvidia

In streaming automatic speech recognition (ASR), it is desirable to reduce latency as much as possible while having minimum impact on recognition accuracy. Although a few existing methods are able to achieve this goal, they are difficult to implement due to their dependency on external alignments. In this paper, we propose a simple way to penalize symbol delay in transducer model, so that we can balance the trade-off between symbol delay and accuracy for streaming models without external alignments. Specifically, our method adds a small constant times (T/2 - t), where T is the number of frames and t is the current frame, to all the non-blank log-probabilities (after normalization) that are fed into the two dimensional transducer recursion. For both streaming Conformer models and unidirectional long short-term memory (LSTM) models, experimental results show that it can significantly reduce the symbol delay with an acceptable performance degradation. Our method achieves similar delay-accuracy trade-off to the previously published FastEmit, but we believe our method is preferable because it has a better justification: it is equivalent to penalizing the average symbol delay. Our work is open-sourced and publicly available (https://github.com/k2-fsa/k2).

Liyong Guo, Xiaoyu Yang, Quandong Wang et al. · nvidia

Knowledge distillation(KD) is a common approach to improve model performance in automatic speech recognition (ASR), where a student model is trained to imitate the output behaviour of a teacher model. However, traditional KD methods suffer from teacher label storage issue, especially when the training corpora are large. Although on-the-fly teacher label generation tackles this issue, the training speed is significantly slower as the teacher model has to be evaluated every batch. In this paper, we reformulate the generation of teacher label as a codec problem. We propose a novel Multi-codebook Vector Quantization (MVQ) approach that compresses teacher embeddings to codebook indexes (CI). Based on this, a KD training framework (MVQ-KD) is proposed where a student model predicts the CI generated from the embeddings of a self-supervised pre-trained teacher model. Experiments on the LibriSpeech clean-100 hour show that MVQ-KD framework achieves comparable performance as traditional KD methods (l1, l2), while requiring 256 times less storage. When the full LibriSpeech dataset is used, MVQ-KD framework results in 13.8% and 8.2% relative word error rate reductions (WERRs) for non -streaming transducer on test-clean and test-other and 4.0% and 4.9% for streaming transducer. The implementation of this work is already released as a part of the open-source project icefall.

Zengwei Yao, Liyong Guo, Xiaoyu Yang et al.

The Conformer has become the most popular encoder model for automatic speech recognition (ASR). It adds convolution modules to a transformer to learn both local and global dependencies. In this work we describe a faster, more memory-efficient, and better-performing transformer, called Zipformer. Modeling changes include: 1) a U-Net-like encoder structure where middle stacks operate at lower frame rates; 2) reorganized block structure with more modules, within which we re-use attention weights for efficiency; 3) a modified form of LayerNorm called BiasNorm allows us to retain some length information; 4) new activation functions SwooshR and SwooshL work better than Swish. We also propose a new optimizer, called ScaledAdam, which scales the update by each tensor's current scale to keep the relative change about the same, and also explictly learns the parameter scale. It achieves faster convergence and better performance than Adam. Extensive experiments on LibriSpeech, Aishell-1, and WenetSpeech datasets demonstrate the effectiveness of our proposed Zipformer over other state-of-the-art ASR models. Our code is publicly available at https://github.com/k2-fsa/icefall.

Fangjun Kuang, Liyong Guo, Wei Kang et al.

The RNN-Transducer (RNN-T) framework for speech recognition has been growing in popularity, particularly for deployed real-time ASR systems, because it combines high accuracy with naturally streaming recognition. One of the drawbacks of RNN-T is that its loss function is relatively slow to compute, and can use a lot of memory. Excessive GPU memory usage can make it impractical to use RNN-T loss in cases where the vocabulary size is large: for example, for Chinese character-based ASR. We introduce a method for faster and more memory-efficient RNN-T loss computation. We first obtain pruning bounds for the RNN-T recursion using a simple joiner network that is linear in the encoder and decoder embeddings; we can evaluate this without using much memory. We then use those pruning bounds to evaluate the full, non-linear joiner network.

Xiaoyu Yang, Wei Kang, Zengwei Yao et al.

Prompts are crucial to large language models as they provide context information such as topic or logical relationships. Inspired by this, we propose PromptASR, a framework that integrates prompts in end-to-end automatic speech recognition (E2E ASR) systems to achieve contextualized ASR with controllable style of transcriptions. Specifically, a dedicated text encoder encodes the text prompts and the encodings are injected into the speech encoder by cross-attending the features from two modalities. When using the ground truth text from preceding utterances as content prompt, the proposed system achieves 21.9% and 6.8% relative word error rate reductions on a book reading dataset and an in-house dataset compared to a baseline ASR system. The system can also take word-level biasing lists as prompt to improve recognition accuracy on rare words. An additional style prompt can be given to the text encoder and guide the ASR system to output different styles of transcriptions. The code is available at icefall.

Zengrui Jin, Yifan Yang, Mohan Shi et al.

The evolving speech processing landscape is increasingly focused on complex scenarios like meetings or cocktail parties with multiple simultaneous speakers and far-field conditions. Existing methodologies for addressing these challenges fall into two categories: multi-channel and single-channel solutions. Single-channel approaches, notable for their generality and convenience, do not require specific information about microphone arrays. This paper presents a large-scale far-field overlapping speech dataset, crafted to advance research in speech separation, recognition, and speaker diarization. This dataset is a critical resource for decoding ``Who said What and When'' in multi-talker, reverberant environments, a daunting challenge in the field. Additionally, we introduce a pipeline system encompassing speech separation, recognition, and diarization as a foundational benchmark. Evaluations on the WHAMR! dataset validate the broad applicability of the proposed data.

Jiatong Zhang, Zengwei Yao, Fanglin Chen et al.

Vision Transformers have achieved remarkable progresses, among which Swin Transformer has demonstrated the tremendous potential of Transformer for vision tasks. It surmounts the key challenge of high computational complexity by performing local self-attention within shifted windows. In this work we propose the Semantic-Aware Local-Global Vision Transformer (SALG), to further investigate two potential improvements towards Swin Transformer. First, unlike Swin Transformer that performs uniform partition to produce equal size of regular windows for local self-attention, our SALG performs semantic segmentation in an unsupervised way to explore the underlying semantic priors in the image. As a result, each segmented region can correspond to a semantically meaningful part in the image, potentially leading to more effective features within each of segmented regions. Second, instead of only performing local self-attention within local windows as Swin Transformer does, the proposed SALG performs both 1) local intra-region self-attention for learning fine-grained features within each region and 2) global inter-region feature propagation for modeling global dependencies among all regions. Consequently, our model is able to obtain the global view when learning features for each token, which is the essential advantage of Transformer. Owing to the explicit modeling of the semantic priors and the proposed local-global modeling mechanism, our SALG is particularly advantageous for small-scale models when the modeling capacity is not sufficient for other models to learn semantics implicitly. Extensive experiments across various vision tasks demonstrates the merit of our model over other vision Transformers, especially in the small-scale modeling scenarios.

Han Zhu, Wei Kang, Liyong Guo et al.

Generating spoken dialogue is more challenging than monologue text-to-speech (TTS) due to the need for realistic turn-taking and distinct speaker timbres. Existing spoken dialogue generation models, being auto-regressive, suffer from slow and unstable inference. To overcome these limitations, we introduce ZipVoice-Dialog, a non-autoregressive zero-shot spoken dialogue generation model built upon flow matching. Key designs include: 1) speaker-turn embeddings for precise speaker turn-taking; 2) a curriculum learning strategy for stable speech-text alignment; 3) specialized strategies to enable stereo dialogue generation. Additionally, recognizing the lack of open-source large-scale spoken dialogue datasets, we curated OpenDialog, a 6.8k-hour spoken dialogue dataset from in-the-wild speech data. Furthermore, we established a benchmark to comprehensively evaluate various models. Experimental results demonstrate that ZipVoice-Dialog achieves superior performance in intelligibility, speaker turn-taking accuracy, speaker similarity, and inference speed. Our codes, model checkpoints, demo samples, and the OpenDialog dataset are all publicly available at https://github.com/k2-fsa/ZipVoice.

Yifan Yang, Xiaoyu Yang, Liyong Guo et al.

Neural Transducer and connectionist temporal classification (CTC) are popular end-to-end automatic speech recognition systems. Due to their frame-synchronous design, blank symbols are introduced to address the length mismatch between acoustic frames and output tokens, which might bring redundant computation. Previous studies managed to accelerate the training and inference of neural Transducers by discarding frames based on the blank symbols predicted by a co-trained CTC. However, there is no guarantee that the co-trained CTC can maximize the ratio of blank symbols. This paper proposes two novel regularization methods to explicitly encourage more blanks by constraining the self-loop of non-blank symbols in the CTC. It is interesting to find that the frame reduction ratio of the neural Transducer can approach the theoretical boundary. Experiments on LibriSpeech corpus show that our proposed method accelerates the inference of neural Transducer by 4 times without sacrificing performance. Our work is open-sourced and publicly available https://github.com/k2-fsa/icefall.

Zengwei Yao, Wenjie Pei, Fanglin Chen et al.

The crux of single-channel speech separation is how to encode the mixture of signals into such a latent embedding space that the signals from different speakers can be precisely separated. Existing methods for speech separation either transform the speech signals into frequency domain to perform separation or seek to learn a separable embedding space by constructing a latent domain based on convolutional filters. While the latter type of methods learning an embedding space achieves substantial improvement for speech separation, we argue that the embedding space defined by only one latent domain does not suffice to provide a thoroughly separable encoding space for speech separation. In this paper, we propose the Stepwise-Refining Speech Separation Network (SRSSN), which follows a coarse-to-fine separation framework. It first learns a 1-order latent domain to define an encoding space and thereby performs a rough separation in the coarse phase. Then the proposed SRSSN learns a new latent domain along each basis function of the existing latent domain to obtain a high-order latent domain in the refining phase, which enables our model to perform a refining separation to achieve a more precise speech separation. We demonstrate the effectiveness of our SRSSN by conducting extensive experiments, including speech separation in a clean (noise-free) setting on WSJ0-2/3mix datasets as well as in noisy/reverberant settings on WHAM!/WHAMR! datasets. Furthermore, we also perform experiments of speech recognition on separated speech signals by our model to evaluate the performance of speech separation indirectly.