Speaker Recognition Robustness to Voice Conversion

Abstract

Security systems relying on voice identification can be threatened by human voice imitation or synthetic voices. As voice conversion can be seen as a sort of voice imitation, this paper analyses the performance of an automatic speaker identification system by using converted voices in order to know how vulnerable such systems are to this kind of disguise. The experiments are conducted by using intra-gender and cross-gender conversions between two males and two females. The results show that, in general terms, the system is more robust to intra-gender converted voices than to cross-gender ones.