A case study for the application of text-independent forensic speaker recognition using Bayesian interpretation

In this study, a Bayesian interpretation framework for forensic automatic speaker recognition is applied. The Bayesian approach is applied to a real world forensic case in which the reference and test utterances are recorded by the police criminology department. Models for accused person and additional 10 individuals unrelated with the case are modelled by adapting each from a universal background model. Gaussian mixture model is used and maximum likelihood linear regression method is applied to adapt each person by using a limited amount of data. The results have shown that the likelihood ratio calculated from the reference and test data seems to be an auxiliary evidence which contributes to the final decision.