Evaluation of two-input speech dereverberation techniques

In this paper we review the acoustical and statistical principles underlying the operation of a two-input signal processing speech dereverberation technique (J.Allen et al., 1977) and describe modifications to this technique based upon experimental and analytical studies of this process. These experimental modifications include selectable short-time binaural coherence estimation techniques, frequency domain smoothing of coherence measures on a critical-band basis, and suppression rules based on thresholded coherence estimates. The effect of these modifications and choices of processing parameters will be discussed in relation to changes in spectrograms and the quality and intelligibility of processed signals.