Neural network for demixing super-Gaussian signals

A new method for separating linear mixtures of statistically independent si gnals with super-Gaussian probability distributions, using a simple neural network, is proposed. The procedure is based on geometric properties, and i t is shown that the maxima of the mixed density distribution belong to stra ight lines, the direction vectors of which, when taken as columns of a matr ix, comprise a demixing matrix. The results obtained with synthetic mixture s of real speech signals are shown.