Blind source separation with low frequency compensation for convolutive mixtures

This paper addresses the blind source separation of convolutive and temporally correlated voice mixtures. We combine natural gradient algorithm and temporal complexity algorithm to preserve the temporal and frequency structures of the original signals. Due to the underlying scaling constraint of natural gradient algorithm, the low frequency components of the original sources are suppressed in the output signals. To compensate for low frequency loss, we use a measure of temporal complexity to recover the low frequency components of the source signals. Simulation results show that the proposed algorithm can well preserve the structure of the original signals both in time and frequency domains.