CONVOLUTIVE DECORRELATION PROCEDURES FOR BLIND SOURCE SEPARATION

R. Vollgraf 1 , M. Stetter 1 , and K. Obermayer 1
D­10587 Berlin, Germany; email: vro@cs.tu­berlin.de

Convolutive decorrelation algorithms form a class of powerful algorithms for blind source separation. In contrast to ICA, they are based on vanishing second order cross correlation functions between sources. We provide an analyze an unifying approach for convolu­ tive decorrelation procedures. The convolutive decor­ relation procedures impose the problem of simultane­ ously diagonalizing a number of covariance matrices. We examine different cost functions for simultaneous diagonalization with respect to the demixing matrix. It turns out, that best performance is achieved for a cost function, that takes the squared sum of the off di­ agonal elements after the diagonal elements were nor­ malized to unity. We then provide criteria for convolu­ tion kernels, that are optimal for noise robustness and which can guarantee positive definite covariance matri­ ces, which are important for reliable convergence.