Abstract

 This paper describes the theory and implementation of dynamic Bayesian networks in the context of speaker identification. Dynamic Bayesian networks provide a succinct and expressive graphical language for factoring joint probability distributions, and we begin by presenting the structures that are appropriate for doing speaker identification in clean and noisy environments. This approach is notable because it expresses an identification system using only the concepts of random variables and conditional probabilities. We present illustrative experiments in both clean and noisy environments and our experiments show that this new approach is very promising in the field of speaker identification.