ODYSSEY 2004 - The Speaker and Language Recognition Workshop
May 31 - June 3, 2004
This paper summarizes the different algorithms developed in ATVS-UPM in order to submit a reliable Likelihood Ratio based forensic system, fully compliant with the bayesian framework for the analysis of forensic evidences, to 2003 NFI-TNO Forensic Speaker Recognition Evaluation. Once identified the main causes and consequences of the erratic estimation of Likelihood Ratios due to forensic conditions, mainly lack of data and mismatch between suspect and questioned speech, several algorithms are proposed and assessed using Switchboard data. Moreover, a new algorithm, TDLRA (Target Dependent Likelihood Ratio Alignment), guarantees efficiently the presumption of innocence for non-target speakers, which is a mandatory condition of any forensic system. The LR-based submitted system is then assessed with NFI-TNO forensic field data, showing an excellent performance in all evaluation conditions, preserving the presumption of innocence and providing a meaningful Likelihood Ratio for any questioned-speech/suspect-speech pair of the evaluation, which could be directly used for reporting to Court under this bayesian forensic framework.
Bibliographic reference. Gonzalez-Rodriguez, Joaquin / Ramos-Castro, Daniel / Garcia-Gomar, Marta / Ortega-Garcia, Javier (2004): "On robust estimation of likelihood ratios: the ATVS-UPM system at 2003 NFI/TNO forensic evaluation", In ODYS-2004, 83-90.