A dual-branch deep interaction network for multi-channel speech enhancement

Abstract

Multi-channel speech enhancement removes noise and reverberation interference from noisy speech signals captured by microphone arrays. In this paper, we propose a dual-branch deep interaction network (DBDINet) for multi-channel speech enhancement, which complements the important features of both time domain and time–frequency domain in the speech signal. We design a waveform and complex spectrum interaction module (WCIM) to interact deeply with the information of two domains and propose an efficient Conformer (eConformer) as a transition layer of the network to improve network efficiency. We conducted extensive experiments on the synthetic AISHELL-1 dataset and the CHiME-3 dataset. The experimental results show that the proposed method achieves competitive performance on several metrics while maintaining lower computational complexity with faster inference speed than existing advanced methods.