Realistic answer verification: An analysis of user errors in a sentence-repetition task

Speech authentication protocols should have a challenge/response feature to be protected against replay attacks. As a result, they need to verify whether the user responded to an interactive prompt. However, it is usually assumed that the user will provide their answer perfectly. In this paper, we report on an ecologically valid user study that we conducted to test this assumption. Our results show that 40% of user answers are imperfect, even in a task as simple as sentence repetition. Error analysis reveals that 60% of the imperfect answers contain small errors that should be deemed acceptable, which increases the total acceptance rate of this task to 84%. We also tested a forced alignment algorithm as a means of verifying answers automatically.