PRISM: A statistical modeling framework for text-independent speaker verification

This paper presents a statistical modeling framework termed as PRISM for text-independent speaker verification. We decompose the verification task into three subtasks: PRobability density estimation, Information metric and Subspace/Manifold learning (PRISM). Subsequently, we take advantages of variational maximum likelihood estimation, Fisher information metric and discriminant locality preserving projection to realize a verification system based on the PRISM framework. We also demonstrate that many current algorithms fall into the PRISM framework and forecast several novel algorithms. Experimental results on the telephone-telephone-English task of NIST SRE 2008 further prove the correctness of the proposed framework.