Multiple TDOA estimation by using a state coherence transform for solving the permutation problem in frequency-domain BSS

Abstract

A novel method to solve the permutation problem for Blind Source Separation (BSS) is presented. According to the acoustic propagation model, in frequency-domain, each separation matrix can be represented with a set of states associated with each source. We formulate a novel transform of the states which is independent of the aliasing and of the permutations since states belonging to all the sources are exploited at the same time. The estimated TDOAs are used to model the propagation of the acoustic wave and to cluster all the frequency components associated to the same source. Experimental results show that the novel approach can be applied to localize and separate sources in challenging situations: two sources have been separated estimating long demixing filters (0.25-0.5s) using widely spaced microphones (0.25 m) in reverberant environment (T60 = 700 ms) and using very short signals (0.5-1 s).