Recognition error measurements from parameterized distance distributions

Formulations are obtained to estimate or predict the performance of a recognition system from parametric descriptions of the statistical distributions of distance measurements output by the recognizer. The dependence of performance estimates on the statistical parameters is shown, and, through this dependence, how performance in one decision mode, verification, is related to performance in another mode, identification. The validity of the formulations and their predictive capability are examined by means of a large experimental data base of distance samples obtained from a template based speaker recognition system. Error rates are tabulated directly from the distances and compared with performance predictions calculated from the formulas. It is shown that verification performance estimates agree well with direct error rate tabulations, while identification performance estimates are subject to a bias attributable to assumptions made in deriving the formulas. The extent to which the outcome of an identification experiment can be used to predict the outcome of a verification experiment, and vice-versa, is discussed.