Regularized subspace-based acoustic multichannel equalization for speech dereverberation

Acoustic multichannel equalization techniques aim at designing robust reshaping filters which reduce the reverberant energy and preserve the perceptual speech quality. Although the recently proposed channel shortening (CS) technique can achieve reverberant energy suppression, its optimization criterion results in a subspace of solutions where each solution yields a different perceptual speech quality. In this paper, we propose a perceptually advantageous subspace-based equalization technique, where the reshaping filter is constrained to lie in the subspace of the multiple CS solutions. Since it has been shown that regularization increases the robustness of equalization techniques to errors in the estimated room impulse responses (RIRs), a regularization term is incorporated in order to control the energy of the reshaping filter. Simulation results for erroneously estimated RIRs demonstrate the advantages of the proposed approach.