Linguistic-Acoustic Forensic Speaker Identification with Likelihood Ratios from a Multivariate Hierarchical Random Effects Model - A Non-Idiot's Bayes' Approach

Abstract

The discriminant performance of a likelihood ratio based on a two-level multivariate model is examined on the speech of 60 male Japanese speakers using non-contemporaneous telephone recordings over uncontrolled channels. The performance is determined for both F-pattern centre frequencies and LPC cepstral coefficients, extracted from three very different phonetic segments only: a vowel, a voiceless fricative and a nasal. The Multivariate Likelihood Ratio is shown to perform well in discriminating samefrom different–speaker pairs, yielding strength of evidence that can be characterised as moderate for F-pattern and, at the least, very strong for the cepstrum. Comparison is made using the same data analysed with a so-called “independence” or “Idiot’s Bayes” LR approach, which ignores correlation between variables. It is shown that, as commonly found, the Idiot’s Bayes approach outperforms the MVLR. The consequences of this finding for forensic speaker identification are alluded to.