Late reverberation suppression based on multi-channel and multi-frame spectral enhancement

Spectral enhancement methods have been widely used for noise and reverberation suppression, which however most of time result in speech distortions. As demonstrated in recent works for noise suppression, the performance in terms of the trade-off between noise reduction and speech distortion can be improved by exploiting the inter-frame correlation. In this paper, we propose to incorporate the inter-frame correlation into reverberation suppression, based on the assumption that late reverberation is uncorrelated to the desired speech signal. Firstly, we present the inter-frame correlation model of the multi-channel noisy-reverberant signals, which are separated into the components of the desired signal, uncorrelated late reverberation and noise. Secondly, we discuss the multi-channel filters for reverberation suppression by exploiting the inter-frame correlation. The power spectral density of late reverberation and noise are estimated separately to design the final suppression filter. As shown in the experimental results, the proposed method can reduce the speech distortion while maintaining robust performance in terms of reverberation suppression in the reverberant and noisy scenarios.