Background Specificity in Forensic Voice Comparison and Its Relation to the Bayesian Prior Probability

This study investigates the effect of background data specifici-ty on likelihood ratio and prior odds, and consequently on the posterior odds outcome. It is motivated by discussions on the correct choice of speaker recognition background, particularly in forensic voice comparison. We performed strictly controlled experiments with the ANDOSL database where background specificity is the sole independent variable. Results show that target and non-target scores are better separated with less spe-cific background, but that in turn priors must be adjusted down. Because the risk of class recognition instead of individ-ual recognition increases with lower background specificity, we suggest that the prior probability in the Bayes formula is factorised into one part that remains in the domain of the trier of fact – as is conventional – and another part that is related to the specificity of the assumed or agreed background.