Solving global permutation ambiguity of time domain BSS using speaker specific features of speech signals

Multidimensional localization of multiple sources using BSS based TDOA estimators, requires the solution of global permutation ambiguity before fusing several TDOA estimations. Since the separation quality of BSS isn´t always perfect, it is not easy to decide which TDOA belongs to which source. Here we study the possibility of using several speaker specific features of speech signal in order to recognize perceptually dominant sources in each one of moderately separated outputs of BSS algorithm. We compare the feasibility of different features in terms of validity rate of decisions and computational complexity.