Examining long-term formant distributions as a discriminant in forensic speaker comparisons under a likelihood ratio framework

This study investigates the use of long-term formant distributions (LTFDs) as a discriminant in forensic speaker comparisons. LTFDs are the distributions calculated for all values of each formant for a speaker in a single recording. Spontaneous speech recordings from 100 male speakers of Southern Standard British English were analyzed from the DyViS Database (Nolan et al. 2009). The recordings were auto-segmented to obtain a minimum of 50 seconds of vowels per speaker. The iCAbS (iterative cepstral analysis by synthesis) formant tracker was used to automatically extract and measure F1-F4. To assess the evidential value of the LTFDs, likelihood ratios (LRs) were computed using a Multivariate Kernel-Density approach (MVKD). It was found that LTFD performs well overall, but much better with different speaker comparisons than same speaker comparisons (97.4% compared to 84% of comparisons providing correct support). LTFD appears to be a good discriminant to include in forensic speaker comparison analyses and offers the added benefit of avoiding potential correlation problems between vowel phonemes. The MVKD based LR results from this study were also found to be comparable to those results in French et al. (2012) and Becker et al. (2008), which used a Gaussian Mixture Model-Universal Background Model LR approach.