Blind Source Separation of Speech Mixtures using a Simple and Computationally Efficient Time-Frequency Approach

A very simple and extremely computationally efficient algorithm for blind separation of two speech sources from two mixtures is presented in this paper. The algorithm exploits the approximate W-disjoint orthogonality of speech signals and assumes specific sensors (microphones) setting that allows the sources to possess a feature we call cross high-low diversity. Two sources are said to be cross high-low diverse (CH-LD) if the two sources are not both close to the same sensor. A source is said to be close to a sensor, if its energy at that sensor is higher than its energy at the other sensor. With this assumption and the W-disjoint orthogonality, it was found that a speech source can easily be extracted from any of the two mixtures with good SIRs (signal-to-interference ratios) based on simple algorithm that compares the ratios of the magnitudes of the time-frequency representations of the two mixtures. The proposed algorithm was tested using different mixtures and has proved to be efficient with both instantaneous and echoic real mixtures. Finally, performance optimization and future expendability to non-CH-LD sources was found possible.