Independent Vector Analysis for Blind Speech Separation Using Complex Generalized Gaussian Mixture Model with Weighted Variance

In this paper, we propose using complex generalized Gaussian mixture distribution with weighted variance for speech modelling and devise an improved independent vector analysis (IVA) algorithm for blind speech separation (BSS). Capable of capturing both non-Gaussianity and non-stationarity, the proposed complex generalized Gaussian mixture model (CGGMM) allows for a much flexible characterization of practical speech signals. The majorization minimization (MM) framework is adopted for the IVA algorithm design. Each iteration of the algorithm is comprised of the updates of demixing matrices and mixture model parameters. For demixing matrices, the update operates in a manner similar to that of the auxiliary function based IVA (AuxIVA) method, and for mixture parameters, the expectation maximization (EM) update is performed. As both updates are in closed form and pre-whitening is not a prerequisite, the IVA algorithm under CGGMM is of low complexity and can be carried out efficiently. Experimental results show that the proposed algorithm outperforms existing ones in terms of separation accuracy and also enjoys a fast convergence rate in both simulated and real environments.