Blind separation of non stationary non Gaussian sources

Most blind sources separation methods are based on the non Gaussianity or the coloration of the sources and only recently their non-stationarity. This work proposes new procedures which exploit both the first and last aspects. We adopt the quasi-maximum likelihood approach which provided a set of estimating equations involving the score functions, which are then estimated by a projection method and through the idea blocking or kernel smoothing. Efficient off-line and on-line algorithms are developed. A simpler and less costly procedure based on a simple contrast for sub Gaussian sources is also considered. Some simulation experiments are given illustrating the high performance of the method.