|
|
Effect of background population on Bayesian conclusions in speaker comparison
Štrosová, Veronika ; Skarnitzl, Radek (advisor) ; Bořil, Tomáš (referee)
Salient goal of forensic phonetics is evaluation of evidence. For this purpose, this field widely uses the Bayesian approach to evidence evaluation, specifically the likelihood ratio (LR). In this probabilistic method of evidence evaluation, which works with concepts of similarity and probability, is background population one of its main cornerstones. This thesis focuses on the effect of background population on the performance of LR based forensic speaker comparison. The LR based classification used midpoint formant values F1, F2 and F3 from Czech vowels of male native speakers. Using Lindleys (1977) parametric LR approach, these acoustic parameters were used to observe the effect of background population varying from 1 to 30 speakers on the validity and reliability of the classification. The analysis showed effect of background population on system performance, whereby the LR output had the lowest validity and reliability using small sample sizes of reference speakers. Simultaneously, the analysis showed an onset of stabilization of the system performance from the perspective of validity as well as reliability from a threshold of background population of 10 reference speakers. In particular cases was also observed synonymous classification validity using very low reference samples with...
|
guest ::
