Fangyuan Wang

Fangyuan Wang, Ming Hao, Yuhai Shi et al.

The conventional recipe for Automatic Speech Recognition (ASR) models is to 1) train multiple checkpoints on a training set while relying on a validation set to prevent overfitting using early stopping and 2) average several last checkpoints or that of the lowest validation losses to obtain the final model. In this paper, we rethink and update the early stopping and checkpoint averaging from the perspective of the bias-variance tradeoff. Theoretically, the bias and variance represent the fitness and variability of a model and the tradeoff of them determines the overall generalization error. But, it's impractical to evaluate them precisely. As an alternative, we take the training loss and validation loss as proxies of bias and variance and guide the early stopping and checkpoint averaging using their tradeoff, namely an Approximated Bias-Variance Tradeoff (ApproBiVT). When evaluating with advanced ASR models, our recipe provides 2.5%-3.7% and 3.1%-4.6% CER reduction on the AISHELL-1 and AISHELL-2, respectively.

Fangyuan Wang, Bo Xu, Bo Xu

Currently, the chunk-wise schemes are often used to make Automatic Speech Recognition (ASR) models to support streaming deployment. However, existing approaches are unable to capture the global context, lack support for parallel training, or exhibit quadratic complexity for the computation of multi-head self-attention (MHSA). On the other side, the causal convolution, no future context used, has become the de facto module in streaming Conformer. In this paper, we propose SSCFormer to push the limit of chunk-wise Conformer for streaming ASR using the following two techniques: 1) A novel cross-chunks context generation method, named Sequential Sampling Chunk (SSC) scheme, to re-partition chunks from regular partitioned chunks to facilitate efficient long-term contextual interaction within local chunks. 2)The Chunked Causal Convolution (C2Conv) is designed to concurrently capture the left context and chunk-wise future context. Evaluations on AISHELL-1 show that an End-to-End (E2E) CER 5.33% can achieve, which even outperforms a strong time-restricted baseline U2. Moreover, the chunk-wise MHSA computation in our model enables it to train with a large batch size and perform inference with linear complexity.

Fangyuan Wang, Zhigang Song, Hongchen Jiang et al.

Most of the recent state-of-the-art results for speaker verification are achieved by X-vector and its subsequent variants. In this paper, we propose a new network architecture which aggregates the channel and context interdependence features from multi aspect based on Time Delay Neural Network (TDNN). Firstly, we use the SE-Res2Blocks as in ECAPA-TDNN to explicitly model the channel interdependence to realize adaptive calibration of channel features, and process local context features in a multi-scale way at a more granular level compared with conventional TDNN-based methods. Secondly, we explore to use the encoder structure of Transformer to model the global context interdependence features at an utterance level which can capture better long term temporal characteristics. Before the pooling layer, we aggregate the outputs of SE-Res2Blocks and Transformer encoder to leverage the complementary channel and context interdependence features learned by themself respectively. Finally, instead of performing a single attentive statistics pooling, we also find it beneficial to extend the pooling method in a multi-head way which can discriminate features from multiple aspect. The proposed MACCIF-TDNN architecture can outperform most of the state-of-the-art TDNN-based systems on VoxCeleb1 test sets.