The purpose of the present paper is to introduce a methodology successfully used already in voice pathology detection for its possible adaptation to biometric speaker characterization as well. For such, the behavior of the same GMM classifiers used in the detection of pathology will be exploited. The work will show specific cases derived from running speech typically used in NIST contests against a Universal Background Model built from the population of normophonic subjects in specific vs general evaluation paradigms. Results are contrasted against a set of impostors derived from the same population of normophonic subjects. The relevance of the parameters used in the study will also be discussed
展开▼
