Automatic speaker recognition with crosslanguage speech material

Abstract

Automatic systems for forensic speaker recognition (FASR) claim to be largely independent of language based on the fact that feature vectors are composed of acoustic parameters that are derived from the resonance characteristics of vocal tract cavities. Yet a certain ‘language gap’ may remain which may deteriorate the performance of a system unless properly compensated. This forensic aspect of what may be called cross-language speaker recognition has not yet received due attention. Based on the most common forensic cross language setting, the aim of this study was to assess the effect of language mismatch on the performance of a standard FASR system and compare its magnitude with the effect of other sources of mismatch on the same voice data. Using the automatic system Batvox 3 in an experiment with 75 bilingual speakers of seven languages and four kinds of transmission channels, it can be shown that, if speaker model and reference population are matched in terms of language, the remaining mismatch between speaker model and test sample can be neglected, since equal error rates (EERs) for same-language or cross-language comparisons are approximately the same, ranging from zero to 5.6%. Transmission of the speech data via landline telephone, GSM and, for part of the corpus, VoIP (using Skype) caused EERs to rise by less than 1% on average.