Time-Domain Multi-Channel Speech Separation Network Using Channel-wise and Inter-channel Features

Abstract

Recently, there has been remarkable performance improvements in single-channel speech separation using deep learning. However, most of single-channel speech separation systems suffer from performance degradation in a challenging environment where reverberation or noise are present. Compared to single-channel speech separation, multi-channel speech separation has the advantage of using additional information from multiple microphones. In this paper, we propose time-domain multi-channel speech separation model based on a fully convolutional time-domain audio separation network (Conv-TasNet). We effectively utilize information from all channels by extracting inter-channel features and channel-wise features. Our results show that the proposed model using channel-wise and inter-channel features improve time-domain multi-channel speech separation performance in both reverberant environment and noisy reverberant environment.