On adaptive multichannel dereverberation based on dichotomous coordinate descent and data-reuse techniques

Abstract

Multichannel linear prediction (MCLP) is widely used for speech dereverberation, with recursive least-squares (RLS)-like algorithms commonly applied to update the linear prediction coefficients. However, these algorithms tend to be computationally intensive, making it necessary in practical implementations to reduce complexity while improving numerical robustness for better dereverberation performance. In this paper, we introduce a more efficient MCLP-based adaptive dereverberation method that combines dichotomous coordinate descent (DCD) with a data-reuse (DR) technique. Compared to the traditional RLS-based approach, the proposed method offers two major benefits. First, it significantly lowers computational demands by replacing most multiplications with bitshifts during DCD iterations, making it more suitable for real-world applications. Second, by avoiding the propagation of the inverse covariance matrix via the Riccati equation, the method ensures numerical stability, making it more suitable for processing long-duration speech signals. Additionally, the DR technique improves dereverberation performance by more efficiently utilizing available observed data. Simulation results show that the proposed methods outperform the conventional RLS-based approach in terms of both numerical stability and computational efficiency, while delivering comparable dereverberation performance.