A class of regularized adaptive multichannel equalization algorithms for speech dereverberation

The multiple input/output inverse theorem (MINT) is a typical speech dereverberation technique based on acoustic channel equalization. Although the adaptive MINT (A-MINT) method has been developed to meet the requirements of real-time application systems, its performance often deteriorates due to the effect of the identification errors of room impulse responses (RIRs). In this paper, we present a regularized A-MINT approach to reduce the sensitivity of the adaptive equalizer to the RIR identification errors. According to the regularized MINT criterion, Newton's and quasi-Newton methods are used to establish two adaptive speech dereverberation algorithms, respectively, which obtain faster convergent rate and better dereverberation performance. The effectiveness of the proposed algorithms is validated through numerical experiments with the RIRs measured in real acoustic environments.