Blind partial extraction of instantaneous mixtures of sources using second order statistics

We introduce a criterion for blindly extracting a small subset of most interesting sources in instantaneous mixtures, based on second order statistics. The extracted sources are those for which the time varying spectral density varies the most. The gradient and an approximation to the Hessian of the criterion are derived. Based on the block diagonal form of the approximate Hessian, an algorithm of relaxation type is developed. A simulation example is provided showing the good performance of the algorithm.