On the potential threat of using large speech corpora for impostor selection in speaker verification

In order to evaluate the risk in SV systems one should take into account the possible perpetrator who knows whom he/she is attacking. This paper thus evaluated if a large speech corpus can be used for selecting impostors to use against a speaker verification (SV) system. We tested the possibility by selecting the most similar speakers from a large corpus and then using the recordings in this corpus for impostor attempts against clients in an SV system. Speech samples of the clients uttering words not used in the SV service was used in order to search the large corpus for similar speakers. Recordings of utterances used in the SV service was then collected from these similar speakers and used for impostor attempts. Our conclusion is that this scenario is a threat that needs to be considered by providers of SV systems.