Complex FastIVA: A Robust Maximum Likelihood Approach of MICA for Convolutive BSS

We tackle the frequency-domain blind source separation problem in a way to avoid permutation correction. By exploiting the facts that the frequency components of a signal have some dependency and that the mixing of sources is restricted to each frequency bin, we apply the concept of multidimensional independent component analysis to the problem and propose a new algorithm that separates independent groups of dependent source components. We introduce general entropic contrast functions for this analysis and a corresponding likelihood function with a multidimensional prior that models the dependent frequency components. We assume circularity for the complex variables and derive a fast algorithm by applying Newton’s method learning rule. The algorithm separates mixed sources even in very challenging acoustic settings.