Reverberation has a detrimental effect on speech perception both in terms of quality as well as intelligibility, as late reflections smear temporal and spectral cues. The ideal binary mask, which is an established computational approach to sound separation, was recently extended to remove reverberation. Experiments with both normal hearing and hearing impaired listeners have shown significant intelligibility improvements for reverberant speech processed using such a priori binary masks. The dereverberation problem can thus be formulated as a classification problem, where the desired output is the ideal binary mask. The goal in this approach is to produce a mask that selects the time-frequency regions where the direct energy dominates the energy from the late reflections. In this study, a binaural dereverberation algorithm is proposed which utilizes the binaural cues of interaural time and level differences as features. The algorithm is tested in highly reverberant environments using both simulated and recorded room impulse responses. Evaluations show significant improvements over the unprocessed condition as measured by both a speech quality measure and a speech intelligibility predictor.
Bibliographic reference. Roman, Nicoleta / Mandel, Michael I. (2013): "Classification based binaural dereverberation", In INTERSPEECH-2013, 3249-3253.