The Bernard Data Set as a Reference Distribution for Bayesian Likelihood Ratio-based Forensic Speaker Identification using Formants

This paper examines the use of John Bernard’s data set of male speakers of Australian English as a reference distribution for the practice of Forensic Speaker Identification in Australian contexts. The Bayesian likelihood ratio-based discrimination is performed on recordings of eleven male speakers of Australian English, using the fpatterns of F1, F2, and F3 of their five tense monophthongs /i/, /a/, /o/, / /, and / /. Target values for these formants and vowels of the speakers described as Broad or General by Bernard are used as the reference distribution for the analysis. The performance of the data set is evaluated using a Likelihood Ratio as discriminant function, and also by an examining of the strength of evidence generated. The analysis is performed using a relatively simple model, assuming normality of the reference sample and equal variance. It is shown that the Bernard data function well, yielding strong strength of evidence. The results of the analysis are discussed with regard to the practice of FSI in Australia.