Relaxed Multichannel Least Squares with Constrained Initial Taps for Multichannel Dereverberation

Abstract:
This paper presents a novel algorithm for robust multichannel dereverberation in the presence of system identification errors with the specific aim of avoiding colouration of the equalized signal. Our proposed algorithm is based upon the technique of channel shortening, which targets only the late taps of the room impulse response. Within the framework of the relaxed multichannel least squares (RMCLS) algorithm, we employ partial relaxation of the early taps of the equalized impulse response (EIR) to increase robustness to channel estimation errors, while constraining the initial taps to avoid undesirable colouration of the equalized signal. It is shown through quantitative experimental results that the resultant equalized signal has an overall improved speech quality perception when compared to alternative algorithms. Index Terms — Multichannel equalization, dereverberation, robustness, system identification errors