Empirical validation of likelihood ratio methods – A case study in forensic speaker recognition

Abstract

In recent years there has been an increase in the offer of commercial software tools for forensic evaluation of biometric evidence. The important role these tools are meant to play requires that these are subject to a rigorous and careful validation process. This has spurred guidelines and procedures through which the validation should be processed. In this work we describe the validation procedure of a commercial Likelihood Ratio computation tool while proposing a new approach to performance evaluation. This is done by testing not only the tool alongside a single baseline method but several baseline methods which have previously been validated by the community. The tool is tested with over 90000 speaker recognition scores and its performance is analyzed and compared to well-known methods, it is shown that the use of multiple baseline methods increases the amount of information and context. This improves decision making in the validation process and helps safeguarding against outliers.