Multichannel online speech dereverberation under noisy environments

Abstract

In this paper, we propose a novel online speech dereverberation with multichannel microphone input signals for noisy environments. Unlike conventional dereverberation methods which optimizes the dereverberation filter by noisy microphone input signals, the proposed method optimizes the dereverberation filter by noiseless microphone input signals so as to achieve a good dereverberation filter under noisy environments. Noiseless microphone input signals are estimated by multichannel Wiener filtering which can be interpreted as combination of multichannel beamforming and time-varying singlechannel Wiener filtering. In multichannel Wiener filtering, residual reverberation which cannot be reduced by the time-invariant dereverberation filter is also reduced. Optimization of the parameters are updated by using the expectation-maximization algorithm in an online manner. Experimental results show that the proposed method can reduce reverberation and background noise effectively in an online manner even when microphone input signals are observed under noisy enviornments.