Noisy cyclo-stationary BSS using frequency-domain pseudo-correlation

In this paper, we propose a new approach for blind separation of noisy linear instantaneous mixtures of cyclo-stationary sources using pseudo-correlation matrices in the frequency domain. This approach is an extension of a new method based on spectral decorrelation that we recently proposed and which assumes that all the cyclo-stationary sources and the stationary noise signals are mutually uncorrelated. Contrary to most of noisy BSS algorithms, our approach provides good performance in the determined case even if the noise signals are colored and/or non-Gaussian and of different variances. The simulation results show the much better performance of our approach in comparison to a classical BSS algorithm.