This is an extension of algorithm EFICA, called Block EFICA, for piecewise stationary and non Gaussian signals. In contrast to EFICA and other classical ICA algorithms, the Block EFICA is able to profit from varying distribution of the original signals and also from their varying variance. Therefore, the method is more flexible which is useful when separating real-world signals that usually possess various features not comprised by standard ICA models. Block EFICA theoretically achieves Cramér-Rao bound for the case of constant-variance signals, i.e. signals whose shape of distribution is changing, but the variance remains the same. The only condition is that the score functions of the original signals are known in advance or should be consistently estimated.
- Z. Koldovský, J. Málek, P. Tichavský, Y. Deville, and S. Hosseini, “Blind Separation of Piecewise Stationary NonGaussian Sources”, accepted for publication in Signal Processing (here).
- Z. Koldovský, J. Málek, P. Tichavský, Y. Deville, and S. Hosseini, “Extension of EFICA Algorithm for Blind Separation of Piecewise Stationary Non Gaussian Sources”, to-be published onICASSP 2008, Las Vegas, April 2008. (here)
- Z. Koldovský, J. Málek, P. Tichavský, Y. Deville, and S. Hosseini, “Performance Analysis of Extended EFICA Algorithm”, technical report nr. 2199, ÚTIA, AV ČR, Oct. 2007.(here)
Matlab code: here
September 2008: After some improvements, we’ve changed the name of the method from Extended EFICA to Block EFICA, which is more apposite.
Dataset: A set of Czech language utterances recorded from TV; 16kHz/16bits
Links: BSSGUI package made by Jakub Petkov