DC-TseNet: A dual-channel time-domain speech enhancement network

In this paper, we propose an end-to-end dual-channel time domain speech enhancement approach, named DC-TseNet, for devices with multiple microphones such as mobile phones used in far-filed scenario like teleconferencing. DC-TseNet incorporates a computationally efficient CNN to form a unified encoder-enhancement-decoder structure that learns clean speech directly using multichannel signals. In addition, DC-TseNet is trained from both intra-channel an inter-channel features to express the relevance and difference between the collected signals from the two microphones, which makes sufficient use of spatial information and reduce the influence of recording direction on the signals. The experimental results show that the proposed dual-channel time-domain approach, with more compact model size, significantly outperforms the LSTM-based frequency-domain method. Furthermore, we find that the inter-channel information, especially the difference between two channels, is more important for a better performance gain.