Blind Second-Order Source Extraction of Instantaneous Noisy Mixtures

The problem of blind source extraction (BSE) for noisy measurements is addressed in the domain of second-order statistics using the linear predictor method. By extending the results from the noise-free case, two methods for the noisy case are proposed, whereby, for rigor, the effect of noise is removed from the cost function. The so introduced algorithms are based, respectively, on the minimization of the normalized mean square prediction error (MSPE), and the minimization of MPSE. The analysis of the derived BSE algorithms is supported by simulations