LLS Colloquium: Understanding individual-level performance in forensic voice comparison

On Wednesday 11th May 2022, Justin Lo (UCL) will be presenting on "individual-level performance in forensic voice comparison".

Understanding individual-level performance in forensic voice comparison

Research into the discriminatory potential of features in forensic voice comparison has typically relied on the use of global metrics to assess their overall performance. Analysis of performance on the level of individual speakers, as well as the reasons behind any variation across speakers, can grant us insights into the nature of the errors in the system, but such kind of analysis is so far rarely performed. In this talk, I present a study that explores the relationship between phonetic data and individual discriminatory performance. Using long-term formant distributions (LTFDs) obtained from Canadian English–French bilinguals, both same-language and cross-language comparisons were conducted within the framework of likelihood ratios. In each case, individuals for whom the system performed exceptionally well or poorly were identified and their LTFDs were compared with those of the other speakers. While acoustic analysis shows clear patterns in the LTFD data of speakers with outlying performance in same-language comparisons, the mapping between individual performance and acoustic behaviour becomes more complex in cross-language comparisons. I discuss these findings in the contexts of the language-specific nature of LTFDs and the relationship between performance and phonetic variability.

The talk will take place at 4pm on Zoom, and there will be an opportunity to ask questions at the end. Please sign up to receive a zoom invite at this link. 

Event poster: Understanding individual-level performance in forensic voice comparison