Robust MVDR Beamformer Based on Complex Gaussian Mixture Model With Phase Prior

Abstract

This paper studies a robust beamforming algorithm for microphone array speech enhancement in the presence of speaker interference and background noise. Accurate steering vector estimation is essential for the performance of the beamformer as well as for the successful speech enhancement. Recently, a time-frequency masking technique based on complex Gaussian mixture model (CGMM) was proposed to efficiently estimate the steering vector for beamforming. However, its performance will degrade with observations that contain noise or/and interference only samples due to the inaccuracy of the CGMM parameter estimation. In this paper, a phase prior for a spatial correlation matrix (a CGMM parameter) is proposed to improve the steering vector estimation in the presence of speaker interference and background noise. Computer simulations are conducted to verify the advantages achieved by the proposed phase prior-based beamformer, in comparison with the conventional beamformer and the CGMM-based approach without prior.