Source separation based on second order statistics-an algebraic approach

Two unknown non-white stochastic sources (e.g. speech signals) are dynamically mixed by an unknown multipath channel and subsequently measured by two sensors. The objective is to construct an inverse filter that separates the two signals, based only on their independence. It is known that, under certain conditions, second-order statistics provide sufficient information to identify the filter. In contrast to the usual cost function optimization techniques, we propose an algorithm that computes the filter coefficients algebraically, using linear algebra techniques such as the singular value decomposition.