Voice-based security: identity verification over telephone lines

The authors discuss potential applications of voice-based security and present accuracy data from a series of evaluative experiments with a Bellcore-developed verifier. In addition, they offer suggestions on how verification performance might be further improved, based on data collected using human listeners. Signal detection theory is shown to provide a valuable framework for assessing and optimizing performance of speaker identity verification systems. The operating characteristic derived from the dissimilarity distributions completely describes the decision space for a verification task, and allows simple comparisons among alternative analyses either by a direct comparison of operating characteristics or comparisons among parameters of characteristics such as the equal error rate. The technique is also invaluable for comparing human and machine performance for selected subsets of utterances. Using this technique, some interesting possibilities have been raised for improving the performance of automated systems by including information that is poorly modeled by current systems but nevertheless provides significant identity-bearing information