Speech separation has been successfully applied as a front-end processing module of conversation transcription systems thanks to its ability to handle overlapped speech and its flexibility to combine with downstream tasks such as automatic speech recognition (ASR). However, a speech separation model often introduces target speech distortion, resulting in a sub-optimum word error rate (WER). In this paper, we describe our efforts to improve the performance of a single channel speech separation system. Specifically, we investigate a two-stage training scheme that firstly applies a feature level optimization criterion for pre-training, followed by an ASR-oriented optimization criterion using an end-to-end (E2E) speech recognition model. Meanwhile, to keep the model light-weight, we introduce a modified teacher-student learning technique for model compression. By combining those approaches, we achieve a absolute average WER improvement of 2.70% and 0.77% using models with less than 10M parameters compared with the previous state-of-the-art results on the LibriCSS dataset for utterance-wise evaluation and continuous evaluation, respectively.
Cite as: Wu, J., Chen, Z., Chen, S., Wu, Y., Yoshioka, T., Kanda, N., Liu, S., Li, J. (2021) Investigation of Practical Aspects of Single Channel Speech Separation for ASR. Proc. Interspeech 2021, 3066-3070, doi: 10.21437/Interspeech.2021-921
@inproceedings{wu21f_interspeech,
author={Jian Wu and Zhuo Chen and Sanyuan Chen and Yu Wu and Takuya Yoshioka and Naoyuki Kanda and Shujie Liu and Jinyu Li},
title={{Investigation of Practical Aspects of Single Channel Speech Separation for ASR}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={3066--3070},
doi={10.21437/Interspeech.2021-921}
}