On IIR Filters and Convergence Acceleration for Convolutive Blind Source Separation

Abstract

It is desirable that online configurations of convolutive source separation algorithms present fast convergence. In this paper, we propose two heuristic forms of increasing the convergence speed of a source separation algorithm based on second-order statistics. The first approach consists of using time-varying learning factors, while the second approach employs a recursive estimation of the short-time autocorrelation functions of the outputs. We also verify, through experiments, whether the cost function considered in the derivation of the algorithm yields, in general, good selection of IIR filters to perform the separation.