On Blocking Matrix-Based Dereverberation for Automatic Speech Recognition

Abstract:
We investigate a two-channel reverberation suppression scheme comprising a blocking matrix for estimating late reverberation components by canceling the direct path and early reflections, and spectral enhancement filters for suppressing the late reverberation components. For idealized blocking matrices, we analyze the influence of the length of the blocking matrix filters and the impact of the coherence between the microphones on the resulting estimate. We show that blocking matrices that cancel more than the direct path of the desired signal can be of advantage for robust speech recognition in highly reverberant environments. Index Terms — Dereverberation, reverberation suppression, signal enhancement, distant-talking speech recognition