Multi-microphone speech dereverberation using spatio-temporal averaging

The use of the source-filter speech production model in methods for enhancement of reverberant speech has received considerable attention over the last few years. Furthermore, it has most recently been shown that spatial averaging of the linear prediction (LP) coefficients is required to improve accuracy in implementation of these types of algorithms. In this paper, we suggest and demonstrate experimentally that LP coefficients obtained from spatially averaged multi-channel speech signals achieve an equally satisfactory result. Consequently, we propose a novel multi-channel speech dereverberation approach operating on the LP residual, utilizing a combination of spatial averaging and a new approach based on inter-cycle temporal averaging. Simulation results and informal listening tests indicate an improvement in terms of direct-to-reverberant sound ratio and in perceived quality of the enhanced speech.