Speech dereverberation with multi-channel linear prediction and sparse priors for the desired signal

摘要

The quality of recorded speech signals can be substantially affected by room reverberation. In this paper we focus on a blind method for speech dereverberation based on the multi-channel linear prediction model in the short-time Fourier domain, where the parameters of the model are estimated using a maximum-likelihood procedure. Contrary to the conventional approach, we propose to model the desired speech signal using a general sparse prior that can be represented as a maximization over scaled complex Gaussians. Experimental evaluation, employing a parametric complex generalized Gaussian prior for the desired speech signal, shows that instrumentally predicted speech quality can be improved compared to the conventional approach.