Spherical discriminant analysis in semi-supervised speaker clustering

Abstract

Semi-supervised speaker clustering refers to the use of our prior knowledge of speakers in general to assist the unsupervised speaker clustering process. In the form of an independent training set, the prior knowledge helps us learn a speaker-discriminative feature transformation, a universal speaker prior model, and a discriminative speaker subspace, or equivalently a speaker-discriminative distance metric. The directional scattering patterns of Gaussian mixture model mean supervectors motivate us to perform discriminant analysis on the unit hypersphere rather than in the Euclidean space, which leads to a novel dimensionality reduction technique called spherical discriminant analysis (SDA). Our experiment results show that in the SDA subspace, speaker clustering yields superior performance than that in other reduced-dimensional subspaces (e.g., PCA and LDA).