Fast spatio-temporal decorrelation using FIR filter network with decoupled adaptive step sizes

Abstract

Blind separation of convolutive mixtures has many applications in the areas of acoustics, communications, biomedical signal processing. Spatio-temporal decorrelation is a useful preprocessing technique that can reduce the blind separation problem into a simpler orthogonal matrix estimation problem. However, it is computationally expensive for its large unknown parameter set. Realtime spatio-temporal decorrelation method remains an open issue for years. In this paper, we focus on accelerating the adaptation process in the FIR filter based decorrelation method. Firstly, the update equation is decoupled into spatial decorrelation and temporal decorrelation. Then, their respective convergence conditions are systematically analyzed using eigenvalue decomposing method. Based on it, a fast algorithm for FIR filter network with spatio-temporal decoupled adaptive step-sizes is proposed. Comprehensive simulations are performed on both synthetic data and real world captured sEMG data. We have compared the proposed algorithm with the fixed step size method, IIR filter with natural gradient, and the annealed step size method. The factors including number of iteration, number of channel, computational time, selection of filter order and window size are elaborately compared and discussed. The results verified the reliable performance of our method on accelerating the adaptation process for spatio-temporal decorrelation simultaneously compared to classical methods.