Robust estimation, interpretation and assessment of likelihood ratios in forensic speaker recognition

摘要: In this contribution, the Bayesian framework for interpretation of evidence when applied to forensic speaker recognition is introduced. Different aspects use voice as in court are addressed, well by expert likelihood ratio right way express strength evidence. Details on computation procedures ratios (LR) given, along with assessment tools and methods validate performance these systems. However, due practical scarcity suspect data mismatched conditions between traces reference populations common daily casework, significant errors appear LR estimation if specific robust techniques not applied. Original contributions fully described, including TDLRA (target dependent alignment), oriented guarantee presumption innocence suspected but non-perpetrators speakers. These algorithms assessed extensive Switchboard experiments moreover through blind LR-based submissions both NFI-TNO 2003 Forensic SRE NIST 2004 SRE, where was successfully provided every questioned speech-suspect recording pair respective evaluations.