A Maximum A Posteriori Approach to Multichannel Speech Dereverberation and Denoising

Abstract:
The transmission of speech in hands-free communication systems is usually corrupted by room reverberation and background noise. In this work, we propose a maximum a posteriori (MAP) algorithm for multichannel speech enhancement in the frequency domain. Based on dynamical modeling of the underlying acoustic channels, we formulate a lower bound on the log posterior and iteratively solve the resulting optimization problem using an online realization of the expectation-maximization (EM) algorithm. The MAP approach enables us to incorporate a priori beliefs about the distribution of the speech signal into the objective function. As a result, we obtain an equalizer that utilizes speech and noise covariances in the source signal estimation. Finally, we derive learning rules for these covariances within the EMframework to obtain a MAP algorithm for blind equalization and channel identification (MAP-BENCH) that jointly estimates the channels and the source signal. Our evaluation shows that the proposed method consistently improves the signal quality in the presence of reverberation and noise. Index Terms — Denoising, dereverberation, expectation-maximization algorithm, maximum a posteriori estimation, speech prior.