Accounting for correlation in linguistic-acoustic likelihood ratio-based forensic speaker discrimination

The necessity of taking correlation between variables into account when estimating strength of forensic speaker recognition evidence is argued for. A modest forensic speaker discrimination experiment is described which investigates how well non-contemporaneous speech samples from the same speaker can be discriminated from different-speaker samples using bivariate kernel density likelihood ratios from F2 and F3 of the five monophthongal phonemes of General Australian English, spoken by 11 males. The experiment shows that an approach which takes the correlation of variables into account can yield useful strengths of evidence. It is also pointed out that the results of such tests still require evaluation with the appropriate confidence limits.