Speech Enhancement by Denoising and Dereverberation Using a Generalized Sidelobe Canceller-Based Multichannel Wiener Filter

This paper describes a speech enhancement system based on a multichannel Wiener filter (MWF) embedded in a generalized sidelobe canceller (GSC). Instead of a fixed beamformer, the minimum variance distortionless response (MVDR) beamformer is used in the signal extracting path. Noise and reverberation covariance matrices are estimated for the MVDR beamformer. The noise power spectral density (PSD) is estimated in light of the spherically isotropic noise model and a blocking-based least-squares method. The reverberation covariance matrix is established using a variance normalization delay linear prediction (NDLP) algorithm. To further reduce residual noise at the GSC output, a postfilter is cascaded. We compare the proposed enhancement system with five baseline methods, the integrated sidelobe cancellation and the linear prediction Kalman filter (ISCLP), the two-stage beamforming approach (TSBA), the blocking-based multichannel Wiener filter (BMWF), the GSC beamformer with a delay-and-sum (DS) as its fixed beamformer and a Wiener postfilter (DS-GSC-PF), and the GSC beamformer with a superdirective beamformer as its fixed beamformer and a Wiener postfilter (SB-GSC-PF). The experiment results have demonstrated that the proposed approach outperforms the ISCLP, TSBA, BMWF, DS-GSC-PF, and SB-GSC-PF methods in terms of objective quality evaluations.