BLIND SEPARATION OF MORE SOURCES THAN MIXTURES USING SPARSITY OF THEIR SHORT­TIME FOURIER TRANSFORM

Pau Bofill \Lambda, Michael Zibulevsky y
pau@ac.upc.es, michael@cs.unm.edu

This paper focuses on underdetermined Blind Source Separation, that is, the separation of N sources from M linear mixtures when M ! N . We exploit the spar­ sity of the short­time Fourier transform when applied to music and speech signals. Given the mixing ma­ trix, a sparse representation of the sources is obtained by solving a low­dimensional linear programming (LP) problem for each of the data points independently. For M = 2 we propose a shortest path, closed­form solution to this LP problem that represents each data point as a linear combination of the pair of directions that en­ close it. The mixing matrix can be estimated either by maximizing a likelihood function, using the above LP optimization at the internal step, or with a clustering algorithm that gives a much faster solution. In this work, for M = 2 we use a clustering algorithm based on a triangular potential function which infers both the mixing matrix and the number of sources. Several experiments involving music and speech signals are described, including the separation of six sources from two mixtures.