COMPLETE TRAINING ANALYSIS OF FEEDBACK ARCHITECTURE NETWORKS THAT PERFORM BLIND SOURCE SEPARATION AND DECONVOLUTION

Nikos A. Kanlis, Jonathan Z. Simon, and Shihab A. Shamma
nkanlis,jzsimon,sas@eng.umd.edu

In this paper we address the diĘcult problem of separating multiple speakers in a real-world situation, were the record- ings are not just instantaneous mixtures, but rather mix- tures of ltered versions of the sources. The enhancement over the approaches already presented by other researchers is that our model allows direct-path, zero-delayed versions of all the sources to be present in each one of the mixtures (a diĘcult approach because it introduces recursiveness in the model, but a closer to the reality one). The update rules are all derived in matrix form (suitable for computing environ- ments, e.g Matlab), with special attention to the diagonals of those matrices, in order to avoid \temporal whitening" at the output. Extending those update rules to ones based on \natural" gradient is also addressed.