Rui Cheng(Beijing University of Technology) and Changchun Bao(Beijing University of Technology)
Speech enhancement is an indispensable technology in the field of speech interaction. With the development of microphone array signal processing technology and deep learning, the beamforming combined with neural network has provided a more diverse solution for this field. In this paper, a multi-channel speech enhancement method is proposed, which combines beamforming and post-filtering based on neural network. The spatial features and phase information of target speech are incorporated into the beamforming by neural network, and a neural network based single-channel post-filtering with the phase correction is further combined to improve the performance. The experiments at different signal-to-noise ratio (SNR) levels confirmed that the proposed method results in an obvious improvement on speech quality and intelligibility compared to the reference methods.