Binaural speech enhancement system combining dereverberation and spatial masking-based noise removal for robust speech recognition

摘要

In this paper, we present a binaural speech enhancement system as a preprocessing step for the robust speech recognition. This system employs an existing dereverberation technique followed by a spatial masking-based noise removal algorithm where only signals coming from the desired directions are retained by using a threshold angle. While state-of-the art approaches fix the threshold angle heuristically over all time frames, we propose to consider an adaptive computation where this threshold angle is first learned in several noise-only frames and then updated frame by frame. Speech recognition results in real environment show the effectiveness of the proposed speech enhancement approach.