LR-based forensic comparison under severe test-data scarcity

This study sets out to find the most reliable method for loglikelihood-ratio (LLR) calculation under severe data scarcity, which is typical of forensic voice comparison casework. We compared the performances of three types of speaker modelling, namely a single Gaussian model, Gaussian Mixture Models (GMM) of different complexity, and a Multivariate Kernel Density Model (MVKD), using two and threedimensional formant frequency feature vectors extracted from /iː/ vowels. We varied the number of tokens used in the offender dataset from 2 to 6. We find that calibration of the systems was critical for dependable evaluation with all the systems tested and that the MVKD model outperformed Gaussian models in most cases.