Language Dependency of Speaker Recognition Systems

Abstract

One of many possible biometric identification methods is voice recognition. A voice recording can be compared to a suspect sample to determine whether suspect and perpetrator are the same. Often the necessary materials are not all available in the same language. This aspect called cross-language speaker recognition can make identification much more challenging. It is therefore fundamental to ensure that the systems used, can correctly perform the assessment in cross-language situations. This paper compares two systems used for biometric speaker recognition that both support cross language identification. The first tool is the BatVOX system, made by Agnitio corp. The second system is Nuance Forensics, made by Nuance. Several contradictions concerning language dependence have been seen. In some situations the language match test performs better than the language mismatch test, while this has been seen the other way around as well. In general the Nuance system seems to be slightly better, however nothing can be said about the language dependency of both systems. It is recommended to obtain more data from both languages in order to make a proper comparison.