Interspeech'2005 - Eurospeech
One of the most important problems in frequency domain blind source separation is the inconsistency across frequency in the permutation of the source estimates. In this paper we present a new algorithm that simultaneously diagonalizes the intra-frequency and the inter-frequency correlation matrices of the source estimates. Experimental results, using speech signals, reveal that the algorithm achieves a highly improved alignment of the permutations between neighbor frequency bins.
Bibliographic reference. Robledo-Arnuncio, Enrique / Juang, Biing-Hwang (2005): "Using inter-frequency decorrelation to reduce the permutation inconsistency problem in blind source separation", In INTERSPEECH-2005, 2357-2360.