Coherence-based Dereverberation for Automatic Speech Recognition

The idea of performing dereverberation using a short-time spatial coherence estimate dates back to 1977 [1], when it was proposed to essentially use the magnitude of the coherence as gain for reverberation suppression. Another heuristic method was recently proposed in [2], where a soft threshold function is used to compute a gain from the coherence magnitude, and the parameters of the threshold function are adapted depending on the histogram of the coherence magnitude in each frequency bin. Short-time coherence estimates have also been investigated in the context of beamforming as a so-called postfilter, and solutions for supression of uncorrelated and diffuse noise have been proposed [3]. In this contribution, we focus on methods where, first, the ratio between direct and reverberation signal components (coherent-to-diffuse ratio, CDR) is estimated from a short-time coherence estimate, and filter weights for reverberation suppression are computed from the CDR using, e.g., the Wiener filter or spectral subtraction rule. We compare and illustrate the behavior of a number of different CDR estimators that have been proposed over the past years, and propose a new variant. Finally, we compare the practical effect of the methods by processing reverberated speech and evaluating the recognition accuracy achieved by an automatic speech recognizer with the processed signals.