A Blind Source Separation of instantaneous acoustic mixtures using Natural Gradient Method

A variety of applications concerning communication signal processing involves recovering unobserved signals or “sources” from several observed mixtures, and “cocktail party effect” is a good paradigm related to this process. Given a set of linearly superimposed acoustic signals without knowledge about the sources makes Blind Source Separation (BSS) a very suitable scheme. A more popular approach of BSS, Independent Component Analysis, has been exploited which basically senses the statistical independence of the source signal estimates to achieve separation. A set of interfering signals present in a typical acoustic environment has been instantaneously combined with a pre-determined mixing matrix. A great weight has been given on an excellent rendition of the Infomax technique of Independent Component Analysis (ICA), called the Natural Gradient Method, to employ a cost function that would yield an optimized de-mixing matrix, producing fairly estimated source signals. By varying the learning rate and the score function, a robust performance of the Natural Gradient has been exhibited, maximizing the separation quality, stability and convergence speed.