Towards Goat Detection in Text-Dependent Speaker Verification

We present a method that identifies speakers that are likely to have a high false-reject rate in a text-dependent speaker verification system (“goats”). The method normally uses only the enrollment data to perform this task. We begin with extracting an appropriate feature from each enrollment session. We then rank all the enrollment sessions in the system based on this feature. The lowest-ranking sessions are likely to have a high false-reject rate. We explore several features and show that the 1% lowest-ranking enrollments have a false reject rate of up to 7.8%, compared to our system’s overall rate of 2.0%. Furthermore, when using a single additional verification score from the true speaker for ranking, the false-reject of the 1% lowest-ranking sessions rises up to 33%.