Non-intrusive regularization for least-squares multichannel equalization for speech dereverberation

Abstract

Acoustic multichannel equalization techniques for speech dereverberation are known to be highly sensitive to estimation errors of the room impulse responses. In order to increase robustness, it has been proposed to incorporate regularization. However, the optimal regularization parameter which yields the highest perceptual speech quality has generally been determined intrusively, limiting the practical applicability. In this paper, we propose an automatic non-intrusive procedure for determining the regularization parameter based on the L-curve. Experimental results show that using such an automatic non-intrusive regularization parameter in a recently proposed partial multichannel equalization technique (P-MINT) leads to a very similar performance as using the intrusively determined optimal regularization parameter. Furthermore, it is shown that the automatically regularized P-MINT technique outperforms state-of-the-art multichannel equalization techniques such as channel shortening and relaxed multichannel least-squares, both in terms of reverberant tail suppression and perceptual speech quality.