Time-frequency domain blind source separation - the independence problem and proposed solution

Time-frequency domain blind source separation (BSS) leads to a seldom mentioned but important problem that generally the independence assumption between source signals collapses in the frequency domain due to inadequate samples. It consequently degrades the performance of all the ICA-based BSS methods. In this paper, we set up a criterion on the performance of separation at each frequency bin and propose a recursive algorithm to correct the bin separations which are thought improper. The bin mixtures are separated into the components of the sources as practical instead of the "independent" bins as achieved by the conventional ICA. The signal-to-noise ratio is greatly increased at certain bins, which results in a much better separation.