Using multiple frequency bins for stabilization of FD-ICA algorithms

Abstract

In the frequency domain independent component analysis approaches for audiosources separation, the convolutive mixing problem is replaced by thesolution of several instantaneous mixing problems, one for each frequencybin of the short time Fourier transform. This methodology yields good resultsbut requires the solution of the permutation ambiguity. Moreover, theperformance of the separation algorithms for each bin is not guaranteed tobe equivalent, thus some bins can have worse results than others. In thispaper a technique based on data from multiple bins is proposed to addressthese issues. The use of multiple bin information produces a coupling of theseparation, resulting in more stable separation matrices and reducing the occurrence of permutations, but increasing in computational cost. This can bemitigated by a sub sampling of the multiple bins information. The resultsshow that both approaches are beneficial for the frequency domain ICA approach,producing better separation in terms of objective quality measures.