Group Multi-Scale convolutional Network for Monaural Speech Enhancement in Time-domain

Recent researches show that convolutional neural networks (CNN) can effectively enhance speech signal by mod-eling its long-term dependence in time-domain. However, the unscaled speech sequence length challenges the receptive field of the convolutional speech enhancement system. This paper proposes a plug-and-play bottleneck module named group multi-scale (GMS) module to alleviate the receptive field craving of convolutional neural networks. The GMS module adopts Group-Communication fashion, where each feature group can send messages to both adjacent groups and output features by convolutional encoding. In this way, the series group forms a sub temporal convolutional network (TCN) in a single residual block, bringing several times the receptive field of the standard bottleneck module. Experimental results on TIMIT datasets show that the proposed module achieves 1.2 dB SI -SNR gain in the TasNet framework compared with the baseline Con-TasNet.