The Geometry of the Channel Space in GMM-Based Speaker Recognition

We describe an extension of the joint factor analysis model of speaker and channel variability in which channel supervectors are modeled by mixtures of low-rank Gaussians rather than by a unimodal Gaussian. This version of the joint factor analysis model includes data-driven feature mapping and the standard joint factor analysis models as limiting cases and it enables us to explore a range of possibilities between these two extremes. Our experimental results indicate that unimodal models of relatively high rank perform better than mixture models of lower rank and they confirm the appropriateness of the unimodal assumption in the standard joint factor analysis model