Disyllabic parameterisation of Vietnamese tonal F0 trajectories in likelihood ratio-based forensic voice comparison

Abstract

The paper presents a preliminary investigation of the performance of acoustic-phonetic based forensic voice comparison features derived from tonal fundamental frequency (F0) trajectories parameterised over disyllabic words, rather than individual syllables as is typically done. Its aim was to see whether the disyllabic parameterisation led to any improvements in the validity and reliability of the voice comparison using speech data from 10 native Vietnamese speakers. Polynomial functions were modelled to speakers' raw F0 trajectories and feature vectors constructed from the model coefficients. Likelihood ratios were calculated using the multivariate kernel density likelihood ratio (MVKD-LR) formula and performance assessed using the log-likelihood ratio cost function (Cllr) (measured in bits) and Bayesian credible interval (BCI). The best performing disyllabic system achieved a Cllr of 0.40, representing a 0.15 and 0.25 bit improvement in validity over the syllable based systems.