Collection and analysis of multi-condition audio recordings for forensic automatic speaker recognition

Abstract

The major aim of the project presented here is to compile a corpus from real case recordings to validate more recording conditions and languages under match and mismatch conditions for forensic automatic speaker recognition (FASR). The challenges and limitations of compiling a real case corpus are explained. First results of validation tests are presented for male speakers of German in the match condition [voice message – voice message] as well as in the mismatch condition [voice message – telephone]. Results for the match condition [voice message] are compared to previous findings for the match condition [telephone]. Variations of performance metrics such as Equal Error Rate (EER) and log-likelihoodratio cost (Cllr) are discussed with respect to effects of normalisation and calibration, and patterns of score distributions are analysed using Tippett plots.