Despite its many prima facie attractive properties for Forensic Speaker Recognition, F0 is regarded as having limited forensic value due to its large within-speaker variability. However, its forensic use to date has been limited mostly to its long-term mean and standard deviation. This paper examines the discriminatory potential, within a Likelihood Ratio-based approach, of additional parametric features from the distribution of long-term F0: its skew, kurtosis, modal F0 and modal density. Motivated by the observation that the shape of the long-term F0 distribution shows less within-speaker occasion-to-occasion difference, we report a forensic discrimination experiment with non-contemporaneous speech samples from 201 male Japanese speakers. Using a multivariate Likelihood Ratio as discriminant distance with the six LTF0 distribution parameters, an equal error rate of 10.7% is obtained from 201 target and 80400 non-target trials. We also investigate how the EER degrades as a function of amount of voiced speech.
1 September 2017 in Forensic Science International
This article investigates to what extent and in what ways the size of the background population affects the outcome of likelihood ratio (LR) based forensic voice comparison. While sample size is known to affect the accuracy of statistical modelling, specific effects in the context of forensic voice comparison are not yet understood. Forensic voice comparison analysts need to work with limited data, but it is unclear how this might impact on the LR-based evaluation of evidence. In this article, we report LR-based speaker comparison experiments using variously sized datasets for background population. They use features derived from long term F0 distribution. We examined their performance in terms of accuracy (closeness to the true value) and precision (reproducibility).
26 June 2014 in The International Journal of Speech, Language and the Law
An experiment in forensic text comparison (FTC) within the likelihood ratio (LR) framework is described. The experiment attempts to determine the strength of authorship attribution evidence modelled with N-grams, which is perhaps one of the most basic automatic modelling techniques. The SMS messages of multiple authors selected from the SMS corpus compiled by the National University of Singapore were used for same- and different-author comparisons. I varied the number of words used for the N-gram modelling (200, 1000, 2000 or 3000 words), and then assessed the performance of each set. The performance of the LR-based FTC system was assessed with the log likelihood ratio cost (Cllr). It is shown in this study that N-grams can be employed within an LR framework to discriminate same-author and different-author SMS texts, but a fairly large amount of data are needed to do it well (i.e. to obtain Cllr < 0.75). It is concluded that the LR framework warrants further examination with different features and processing techniques.