Abstract
Multichannel Wiener Filters (MWFs) represent one important class of speech enhancement techniques for denoising and de-reverberation. Nonblocking and blocking approaches are employed to reliably estimate the covariance matrix and power spectral density (PSD) in implementing MWFs. However, for low signal-to-noise ratio (SNR) signals, the blocking-based estimator is prone to artifacts. In this paper, a noise covariance matrix estimation approach is proposed to mitigate speech artifacts frequently encountered in the blocking-based MWF. Noise signals estimated at each channel by a scalar noise Wiener filter. The noise signal estimates of all channels serve to construct the noise covariance matrix for the MWF. A minimum variance distortionless response (MVDR) beamformer and a scalar postfilter are constructed on the basis of the estimated noise and reverberation covariance matrices. This paper provides two methods for estimating the reverberation covariance matrix: a Blocking-Based Noise Signal Estimator (BB-NSE) and a Blocking-Based Noise Signal Estimator and Reverberation Extraction (BB-NSE-R). Simulation results have shown that the BB-NSE has enhanced speech quality satisfactorily in terms of frequency-weighted segmental SNR, signal-to-distortion ratio (SDR) and perceptual evaluation of speech quality (PESQ), as compared to the conventional Blocking-Based MWF (BB).