Multichannel Blind Separation of Sources Algorithm For Instantaneous Mixture

The algorithms of blind separation of sources, in the general case and for instantaneous mixtures, are based on high-order statistics; most of them use the fourth-order statistics. For an instantaneous mixture of only two sources, we proposed in [14] an algorithm of blind separation of sources. The separation was achieved by minimizing the cross-cumulant (2x2) of the two output signals. The minimization of that cross-cumulant was achieved using a gradient algorithm. In this paper, we derive a new cost function which is more general than the first one, also based on the cross-cumulant (2x2) of the output signals. This new algorithm deals with Multiple Inputs and Multiple Outputs (MIMO) and uses a LevenbergMarquardt method for the minimization of the cost function. The actual algorithm is very fast; the criterion convergence is attained in less than 50 iterations. In addition, it yields good results even in the case of about 300 signal samples. Good experimental results were obtained even with five stationary signals.