Graphical models for text-independent speaker verification

Our approach in text independent Speaker Verification (SV) proposes to integrate different aspects of the speech signal which convey information about the speaker's identity using Graphical Models (CM). Prosodic, spectral and source information obtained from the residue of linear prediction analysis are modeled in a probabilistic framework with a system based on Bayesian Networks (BN). The structure, or conditional independencies between the variables, is learned directly from the data using two different algorithms. In particular, the interpretation and comparation of the structures is presented. Some experiments conducted on the NIST 2003 one speaker text-independent data base have been conducted to demonstrate the feasibility of this approach.