Extended Variability Modeling and Unsupervised Adaptation for PLDA Speaker Recognition

Probabilistic Linear Discriminant Analysis (PLDA) continues to be the most effective approach for speaker recognition in the i-vector space. This paper extends the PLDA model to include both enrollment and test cut duration as well as to distinguish between session and channel variability. In addition, we address the task of unsupervised adaptation to unknown new domains in two ways: speaker-dependent PLDA parameters and cohort score normalization using Bayes rule. Experimental results on the NIST SRE16 task show that these principled techniques provide state-of-the-art performance with negligible increase in complexity over a PLDA baseline.