The general source separation problem

Abstract

Presents a maximum likelihood approach of the separation problem of independent sources linearly combined. This separation cannot be successful using only the spectral matrix. But it can be done using higher order statistics and in particular the fourth order cumulant of the signals. The general linear case can be solved in frequency domain. In this domain the authors introduce the notion of circular complex variables, and give the maximum likelihood estimator of the source model in the complex data case.<>