Speaker clustering via novel pseudo-divergence of Gaussian mixture models

Methods based on difference between Gaussian mixture models (GMMs) are widely used in speaker clustering. The paper presents a novel pseudo-divergence, the ratio of inter-model dispersion to intra-model dispersion, to characterize the difference between two GMMs. In dispersion, weight, mean and variance, of which a GMM is composed, are involved. Experiments show that such measurement can well characterize the difference between two GMMs and has good performance in speaker clustering.