On the Performance and Use of Speaker Recognition Systems for Surveillance

We model the performance of a speaker recognition system used for surveillance to prioritize a large number of candidate speakers in search of a single target speaker. It is assumed that the system operates by ordering all speakers in order from best match to worst match, with the goal of having the true speaker sample positioned as high as possible on the list. Some performance measures for prioritization systems are given and are applied to a real speaker recognition system. An analytic expression for the probability density function of the true speaker's position on the list is found, subject to basic assumptions concerning the distribution of true speaker and false speaker scores. A comparison is made to the performance of a system which is operating only by making verification decisions, and it is shown that making soft decisions results in significantly better surveillance performance.