A Soft Decision Based Noise Cross Power Spectral Density Estimation for Two-Microphone Speech Enhancement Systems

Abstract

The coherence function has been successfully applied to two-microphone speech enhancement systems. Recently, cross power spectral density (cross-PSD) estimation was proposed to further improve the coherence based two-microphone speech enhancement systems, especially for the case that the cross correlation between the two channel noises can not be ignored. In this paper, we propose a more robust noise cross-PSD estimation method based on an effective soft decision approach using minimum statistics, which can work well even in highly adverse conditions. Experimental results using real signals recorded in a running car demonstrate the proposed approach can obtain more less speech distortion with the same SNR improvement.