Underdetermined Blind Source Separation using Matching Pursuit Method

Abstract

Based on the sparse characteristics in representation space, we propose and efficient method for separating speech signals. This paper focuses on the sparse distribution in the time-frequency domain. To get mixing matrix used to obtain original signals, we applied STFT to mixtures for making signals sparser. The scatter plot contains information on the sources with their directions. Angles from base line which are represented in scatter plot are used to determine boundaries of segments. We show how to separate the signals by the matching pursuit method and how to identify the moving sources, eliminating the permutation problem. It is concluded that the proposed method is very simple compared to other methods, so this method has great potential for source separation.