Robust likelihood ratio estimation in Bayesian forensic speaker recognition

In this paper we summarize the bayesian methodology for forensic analysis of the evidence in the speaker recognition area. We also describe the procedure to convert any speaker recognition system into a valuable forensic tool according to the bayesian methodology. Furthermore, we study the difference between assessment of speaker recognition technology using DET curves and assessment of forensic systems by means of Tippet plots. Finally, we will show several complete examples of our speaker recognition system in a forensic environment. Some experiments will be presented where, using Ahumada-Gaudí speech data, we optimize the Likelihood Ratio computation procedure in order to be robust to inconsistencies in the estimation of withinand between-sources statistical distributions. Results in the different tested situations, summarized in Tippet plots, show the adequacy of this approach to daily forensic work.