Blind signal separation for convolutive mixing environments using spatial-temporal processing

In this paper we extend the infomax technique of Bell and Sejnowski (see Neural Computation, vol.7, no.6, p.1129-59, 1995) for blind signal separation from the instantaneous mixing case to the convolutive mixing case. Separation in the convolutive case requires an unmixing system which uses present and past values of the observation vector, when the mixing system is causal. Thus, in developing an infomax process, both temporal and spatial dependence of the observations must be considered. We propose a stochastic gradient based structure which accomplishes this task. The performance of the proposed method is verified by subjective listening tests and quantitative measurements.