The effect of correlation on strength of evidence estimates in Forensic Voice Comparison: uni-and multivariate Likelihood Ratio-based discrimination with Australian English vowel acoustics

The consequences of ignoring correlations between features in traditional forensic speaker recognition are investigated. Two likelihood ratio-based discrimination experiments on the same multivariate formant data are described, one taking correlation into account and the other not doing so. The discrimination is performed using Naive Bayes univariate, and multivariate generative Likelihood Ratios (LRs) as discriminant functions, exemplified with Tippett plots and evaluated with the C-llr cost function. It is shown that ignoring within-segment correlation can result in considerable over-or under-estimation of the strength of evidence when traditional features are used, and there is poorer overall discrimination between same-speaker and different-speaker pairs. The use of logistic-regression fusion to handle between-segment correlation is also demonstrated.