Source separation using second order statistics

Abstract

It is often assumed that blind separation of dynamically mixed sources can not be accomplished with second order statistics. In this paper it is shown that separation of dynamically mixed sources indeed can be performed using second order statistics only. Two approaches to achieve this separation are presented. The first approach is to use a new criterion, based on second order statistics. The criterion is used in order to derive a gradient based separation algorithm as well modified Newton separation algorithm. The uniqueness of the solution representing separation is also investigated. The other approach is to use System Identification. In this context system identifiability results are presented. Simulations using both the criterion based approach and a Recursive Prediction Error Method are also presented.