DE-DPCTnet: Deep Encoder Dual-path Convolutional Transformer Network for Multi-channel Speech Separation

Abstract

In recent years, beamforming has been extensively investigated in multi-channel speech separation task. In this paper, we propose a deep encoder dual-path convolutional transformer network (DE-DPCTnet), which directly estimates the beamforming filters for speech separation task in time domain. In order to learn the signal repetitions correctly, nonlinear deep encoder module is proposed to replace the traditional linear one. The improved transformer is also developed by utilizing convolutions to capture long-time speech sequences. The ablation studies demonstrate that the deep encoder and improved transformer indeed benefit the separation performance. The comparisons show that the DE-DPCTnet outperforms the state-of-the-art filter-and-sum network with transform-average-concatenate module (FaSNet-TAC), even with a lower computational complexity.