FORENSIC SPEAKER DISCRIMINATION WITH AUSTRALIAN ENGLISH VOWEL ACOUSTICS

Abstract

A large-scale forensic discrimination experiment is described that investigates how well same-speaker speech samples can be discriminated from different-speaker speech samples using acoustic parameters from Australian English vowels. A multivariate likelihood ratio is used as a discriminant function on the five tense and six lax vowel phonemes of 171 male speakers. In 171 target trials and 58,140 non-target trials, comparing samples with just one token per vowel each gives EERs of between 17% and 40%, which drop to between 10% and 14% when fused. Kernel density modeling of the reference population is shown to outperform normal, and performance degrades under quasi-realistic conditions.