Abstract
Both spatial and temporal contextual information are essential for the multichannel speech enhancement (MCSE) task. In this work, we propose a unified MCSE network composed of neural beamforming and a context-focused post-filtering network in order to fully exploit these two types of information. The network is used to estimate the optimum complex ideal ratio masks (cIRM) which can more effectively utilize the phase information in the frequency domain to reconstruct the speech waveform. To assign adaptive weights to channels, we first adopt a dilated convolution-based network which simulates the beamforming on the original multichannel input spectrum as the front-end of the multichannel acoustic model. Furthermore, we propose a post-filtering network that inputs the suggested U-Net’s output to a convolutional long short-term memory (ConvLSTM) layer, as it can properly capture contextual information and spatial correlation information of features. We conduct experiments on the VOICES, CHiME-3 and WMIR datasets, respectively. Experiments show that, in various scenarios, the proposed algorithm shows improvements over the previous state-of-the-art algorithms in terms of PESQ, STOI and SI-SNR.
Funding Information
  • National Key Research and Development Program of China (2020YFC2004002, 2020YFC2004003)
  • National Natural Science Foundation of China (62001215)
  • Key Research and Development Program of Jiangsu Province (BE2022059-3)