A method of blind separation for convolved non-stationary signals (Q1272398)

The blind separation of sources is a well known method for extracting one or more useful signals which are included in a mixture. The most part of previous papers devoted to this subject assumed the stationarity of the components. The basic assumptions adopted in this paper are: the included signals are connected by convolutive operations (whence the denomination of convolved sigals) and the components are allowed to be nonstationary (i.e. speech, music), this fact being the novelty of the contribution. The proposed method uses two ingredients: an adaptive feedforward network, whose parameters are changed according to a learning rule, based on the natural gradient, and a cost functional which is to be minimized. The proposed method is very interesting. It was validated by computer simulation whose results are presented at the end of the paper.