Blind Adaptive Dereverberation of Speech Signals Using a Microphone Array

This paper describes a blind adaptive method for the dereverberation of speech/audio signals in a closed room environment based on the use of multiple microphones (two or more) by utilizing the second-order statistics of the reverberated speech signals only. The spatial diversity provided by the microphone array creates the equivalent of multiple channels, where each channel is the impulse response of the source audio/speech signal to each microphone. Mathematically, this is equivalent to a single-input multiple-output (SIMO) channel model. The dereverberation is accomplished through an inversion procedure of the channels. The inverse filters, also called equalizers, are found by minimizing what we term a reduced mutually referenced equalizers (RMRE) error criterion. The error criterion is minimized through an iterative procedure that is implemented using linear adaptive filters. The adaptive algorithm used can be one such as the LMS, RLS or any variant of them such as subband implementations. The proposed method was tested on simulated reverberated speech signals with good dereverberation results.