Multi-channel Time-Varying Covariance Matrix Model for Late Reverberation Reduction

Abstract

In this paper, a multi-channel time-varying covariance matrix model for late reverberation reduction is proposed. Reflecting that variance of the late reverberation is time-varying and it depends on past speech source variance, the proposed model is defined as convolution of a speech source variance with a multi-channel time-invariant covariance matrix of late reverberation. The multi-channel time-invariant covariance matrix can be interpreted as a covariance matrix of a multi-channel acoustic transfer function (ATF). An advantageous point of the covariance matrix model against a deterministic ATF model is that the covariance matrix model is robust against fluctuation of the ATF. We propose two covariance matrix models. The first model is a covariance matrix model of late reverberation in the original microphone input signal. The second one is a covariance matrix model of late reverberation in an extended microphone input signal which includes not only current microphone input signal but also past microphone input signal. The second one considers correlation between the current microphone input signal and the past microphone input signal. Experimental results show that the proposed method effectively reduces reverberation especially in a time-varying ATF scenario and the second model is shown to be more effective than the first model.