Underdetermined blind separation of speech signals with delays in different time-frequency domains

Abstract

This paper is devoted to the problem of speech signal separation from a set of observables, when the mixing system is underdetermined and static with unknown delays. The approaches appeared in the literature so far have shown that algorithms based on the property of sparsity of the original signals (effectively satisfied by speech sources) can be successfully applied to such a problem, specially if implemented in the time-frequency domain. Here, a survey on the usage of different time-frequency transforms within the already available three-step procedure for the addressed separation problem is carried out. The novelty of the contribution can be seen from this perspective: Wavelet, Complex Wavelet and Stockwell Transforms are the new transforms used in our problem, in substitution of the usual Short Time Fourier Transform (STFT). Their performances are analyzed and compared to those attainable through the STFT, evaluating how much different is the influence that their sparseness and spectral disjointness properties on the algorithm behavior. © pringer-Verlag Berlin Heidelberg 2005.