A Real-Time Speech Dereverberation Environment Based on Multichannel Parametric Room Equalization

Abstract:
The dereverberation of speech signals is of great importance for hands-free communication systems and automatic speech recognition. Despite a lot of research, the performance of current systems is still disappointing and reverberation remains a challenging and unsolved problem. In this paper, we explore performance limits by online identifying room impulse responses with a known reference signal and by equalizing their resonant and non-resonant characteristics in real-time with multichannel parametric algorithms. The proposed equalizer is highly flexible and particularly designed for the limited resources of a real-time implementation. After a description of our system and its complexity, we evaluate the dereverberation performance of our approach in terms of signal-to-noise ratios and spectrograms.