Blind Dereverberation of Monaural Speech Signals Based on Harmonic Structure

摘要

This paper presents a new method for dereverberating monaural speech signals. Speech signals captured by a distant microphone usually contain a lot of reverberation, which severely degrades the performance of speech applications such as automatic speech recognition (ASR) systems. Although a number of dereverberation methods have been proposed, dereverberation remains a challenging problem, especially when using a single microphone. To overcome this problem, we propose a dereverberation method based on an inherent property of speech signals, namely, their harmonic structure. We show that a filter that enhances the harmonic structure of reverberant speech signals approximates an inverse filter of the reverberation process, and can achieve high-quality speech dereverberation. Experimental results show that a dereverberation filter trained with a sufficient amount of observed reverberant signals can effectively reduce the signal reverberation when the reverberation time is 0.1 to 1.0 s.