Speech Dereverberation Based on Recursive Weighted Prediction Error

摘要

This paper proposes a speech dereverberation method based on recursive weighted prediction error (RWPE) for a moving average (MA) model of reverberation observed by a single distant microphone. We estimate an infinite impulse response (HR) filter for inverse filtering based on a maximum likelihood criterion on a time-varying Gaussian speech model. To make the IIR filter stable, we propose an iterative process of avoiding recursive filtering at the current processing time. We compared the proposed RWPE method with a dereverberation method based on weighted prediction error (WPE) in an application to an in-house automatic speech recognition (ASR) system. The WPE method is known to be an effective method when using multi-channel microphones. However, it mismatches the MA observation model when using a single-channel microphone. Experimental results using simulated reverberant speech reveal that the proposed method achieves a better performance than the WPE method in terms of the accuracy of speech recognition.