A Time-Varying Forgetting Factor-Based QRRLS Algorithm for Multichannel Speech Dereverberation

Abstract

In this paper, we propose an adaptive multichannel linear prediction (MCLP) algorithm based on QR-decomposition recursive least squares (QRRLS) approach for online speech dereverberation, in which a time-varying forgetting factor (VFF) control scheme is devised to adapt to dynamic acoustic scenarios. Being capable of avoiding the numerical instability problem inherent to RLS-based MCLP, QRRLS-based MCLP method shows more robustness while retains the same arithmetical complexity and fast convergence as the RLS-based methods. The VFF scheme based on the approximated derivatives of the filter coefficients is adopted to update the time-wise forgetting factor which can track the varying paths of reflections effectively. Experimental results show that the proposed VFF-QRRLS-based MCLP algorithm improves the performance of speech dereverberation and also enjoys a fast tracking capability and numerical robustness compared with the conventional adaptive MCLP algorithms.