Optimal multichannel equalization for robust speech dereverberation

Ideal acoustic multichannel equalization technique based on multiple-input/output inverse theorem (MINT) aims to completely invert the estimated room impulse responses (RIRs) between the source and microphone arrays. However, the method is known to be very sensitive to the estimation errors of the RIRs. In order to increase robustness, regularized MINT (R-MINT) has been proposed that overcomes the limitation by decreasing the energy of the equalization filters. In this paper, we analyze the solution of the R-MINT method and introduce a number of different ways to increase the robustness of the MINT method. Recently, regularized partial MINT (RP-MINT) method has been proposed that provides better perceptual speech quality compare to the R-MINT method. We also propose a new regularized optimal partial MINT (ROP-MINT) method which aims to optimize the equalized impulse response (EIR) between the source and output of the system. Simulation experiments conducted in various reverberant environments demonstrate that performance of our proposed ROP-MINT is better than that of the RP-MINT method.