Seeing the trees in the forest

Diagnosing individual performance with acoustic data in likelihood ratio based forensic voice comparison

Abstract

System testing is a crucial part of likelihood ratio based forensic voice comparison, but the evaluation of system performance has thus far focused on global metrics, with little attention paid to the variation in individual performance and the factors behind such variation. Using long-term formant distributions as a case study, this study applies the notion of biometric menagerie to analyse performance on the individual level, and further explores the connection between performance and the underlying acoustic data. Zooplot analysis reveals distinct distributions of how individual speakers perform for each long-term formant. Acoustic analysis further reveals clear patterns in the formant data displayed by speakers with outlying performance. Together, the findings support the view that individual-level analysis can offer useful diagnostic insights into system performance that are unavailable in global-level assessment.