Feature selection and model optimization for semi-supervised speaker spotting

摘要

We explore, experimentally, feature selection and optimization of stochastic model parameters for the problem of speaker spotting. Based on an initially identified segment of speech of a speaker, an iterative model refinement method is developed along with a latent variable mixture model so that segments of the same speaker are identified in a long speech record. It is found that a GMM with moderate number of mixtures is better suited for the task than a large number mixture model as used in speaker identification. Similarly, a PCA based low-dimensional projection of MFCC based feature vector provides better performance. We show that about 6 seconds of initially identified speaker data is sufficient to achieve > 90% performance of speaker segment identification.