How to test the performance of speech verifiers and statistical evaluation

Abstract

Biometric identification and verification technologies, in the past, have promised high performance levels. Such performance statements lead to the assumption, that these biometric systems are highly secure. Field data tests have shown substantial discrepancy compared to specified error rates. In order to reflect target scenario deployments when gathering test data, we suggest to acquire test data from actual deployments, which implies a test population reflecting that of the target group. Four impostor levels are defined. For statistical analysis we suggest Sequential Testing according to Wald, in order to minimize population size and still show the statistical significance of low empirical error rates.