Multi-channel Speech Enhancement Based on the MVDR Beamformer and Postfilter

Abstract

Deep neural network (DNN) based ideal ratio mask (IRM) estimation methods have yielded good performance in monaural speech enhancement. Meanwhile, these methods have also shown considerable potential for beamforming and multichannel speech enhancement. It is crucial for minimum variance distortionless response (MVDR) beamformer to estimate the covariance matrix of the speech and noise accurately. The accurate estimation of time-frequency (T-F) mask has significant impact on the estimation of the covariance matrices. So, in this paper, a complex real and imaginary ratio mask (CRIRM) based MVDR beamformer for speech enhancement using residual network is proposed. First, the real and imaginary masks of speech and noise are estimated by taking advantage of a residual neural network. After that, the estimations of speech and noise are obtained by using the estimated masks. Finally, the covariance matrices of speech and noise are estimated, and applied into the MVDR beamformer. In addition, in order to further reduce residual noise interference, the output of the MVDR beamformer is further processed by an end-to-end monaural speech enhancement module. Experiments show that, the proposed method can better improve the quality and intelligibility of the enhanced speech.