Blind source separation using monochannel overcomplete dictionaries

Abstract

We propose a new approach to underdetermined Blind Source Separation (BSS) using sparse decomposition over monochannel dictionary atoms and compare it to multichannel dictionary approaches. We show that the new approach is easily extended to any single channel decomposition method and allows for faster computation of algorithms such as the Bounded Error Subset Selection (BESS) because of the reduced dimension of the search space. Experimental results on Matching Pursuit (MP) and BESS algorithms show that our method can give better Signal to Interference Ratio performance than pursuit methods based on multichannel dictionary atoms.