Blind speech separation using high order statistics

This paper deals with blind speech separation of instantaneous and convolutive mixtures of non-Gaussian sources. The separation criterion is based on higher order statistics (HOS) on the assumption that the sources are statistically independent. We propose to simplify and to improve the classical Herault-Jutten algorithm by choosing adequate high order non-linear functions for adaptation. The convolutive case is investigated through a model with impulse responses modeling the Head Related Transfer Function (HRTF). Experimental results show the efficiency of the proposed approach in terms of signal-to-interference ratio, when compared to the widely used fastICA algorithm. In the convolutive case a satisfactory separation of the sources has been achieved.