A comparative study of algorithms for blind source separation in the instantaneous linear mixture model

Abstract

This thesis discusses some of the many techniques for performing blind source separation. Its focus is on the theoretical concepts that allow for the problem to be solved. It starts with presenting the EM algorithm, which is the method underpinning many of the algorithms that are presented later in the thesis. Some of the established methods are presented, and we proceed to devolop source separation algorithms based upon modelling the sources as scale mixtures of Gaussians. Such models are particularly well suited at modelling the super-Gaussian probability densities that characterise many real world signals, speech being perhaps the most commom.

When evaluating the performance of the algorithms in this thesis, our focus is mainly on the quality of separation, and discussions on computational efficiency are mostly superficial.

We find that in particular one of the algorithms we have constructed shows promise. Its performance is on par with existing methods, and further examination of its properties might be in order.