A distributed decision approach to speaker verification

A speaker-verification system based on an HMM connected word recognizer and a distributed detection decision algorithm is presented. We discuss a text-independent system in which the local decisions are based on individual spectral parameters, and a text-dependent system in which the local decisions are based on the spectral parameters for a single word. The text-independent system assumes the parameter vectors to be normally distributed, while the text-dependent system uses a nonparametric test based solely on the imposter distribution. In both cases, the distributed approach provides robustness against discrepancies between the statistical characteristics of the speech parameters in the training and the test phases. Both distributed-decision speaker verification systems have been evaluated on a database of speech collected during the trial of a real verification system and have been shown to be superior to systems based on conventional decision algorithms