Speaker Identification by BYY Automatic Local Factor Analysis based Three-Level Voting Combination

Abstract

Local Factor Analysis (LFA) is known as more general and powerful than Gaussian Mixture Model (GMM) in unsupervised learning with local subspace structure analysis. In the literature of text-independent speaker identification, GMM has been widely used and investigated, with some preprocessing or postprocessing approaches, while there still lacks efforts on LFA for this task. In pursuit of fast implementation for LFA modeling, this paper focuses on the Bayesian Ying-Yang automatic learning with data smoothing based regularization (BYY-A), which makes automatic model selection during parameter learning. Furthermore for sequence classification, based on trained LFA models, we design and analyze a three-level combination, namely sequence, classifier and committee, respectively. Different combination approaches are designed with variant sequential topologies and voting schemes. Experimental results on the KING speech corpus demonstrate the proposed approaches' effectiveness and potentials.