Blockwised active source number estimation for blind source separation

Abstract

In this paper, a blind source separation (BSS) algorithm that tracks the number of active independent sources in each observation block is considered. Until now, most BSS algorithms assume a fixed number of sources throughout the observations. In interval blocks with some inactive sources, Bayesian based BSS algorithms produce an erroneous outputs. Here, the algorithm estimates the number of active source on a block-by-block bases using reversible-jump Markov chain Monte Carlo (RJ-MCMC), and determines which sources are active. Afterwords, the algorithm splits and concatenates observation blocks where the same sources are active. For each concatenated blocks, separate BSS algorithms are applied, and their outputs are appropriately combined. The considered algorithm does not need to manually select the number of sources. Experimental results show that the proposed algorithm outperform conventional algorithms in BSS.