Blind separation of convolutive mixtures using second and fourth order moments

This paper deals with the problem of blind identification of p-non Gaussian inputs, q-outputs AR systems in the special case where p<q. In this case, the identification problem is degenerated, therefore as the classical Levinson (Robinson) algorithm cannot be applied, we use the Inouye´s (1983) method. As this procedure assumes that the AR model is normalized, it is necessary to split the problem in two parts: first, we estimate the convolutive mixture by means of linear prediction, and second, we estimate the instantaneous mixture. The first one requires second order moments and the second one, high order statistics. Numerical simulations are presented to show the influence of the conditioning of the instantaneous mixture matrix in the identification problem in presence of white Gaussian noise