Abstract
The problem of underdetermined blind source separation is addressed. The sparse assumption which is commonly required in the current underdetermined blind source separation literature is relaxed. By introducing an advanced clustering technique based upon self-splitting competitive learning, the time-frequency plane is partitioned into appropriate blocks where the number of active sources is no more than the number of sensors, resulting in a novel robust block based algorithm. Simulation studies are presented to support the proposed approach for the separation of GMSK sources.