The voice parametric features extraction followed by some statistical identification test is the widespread approach to the speaker recognition. The problem reaches its maximal complexity when a voice sample must be attributed to a set of speakers being non-zero the a-priori probability that the sample does not belong to the given set. Usually these open set tests are related to the highest level of responsibility like in the forensic applications. This paper is addressed to present a balanced solution to the speaker recognition problem and to give the right statistical foundation to the decision task. All the related issues are restated, the modelization method is reconsidered for sparse experimental matrices and the algorithms for a suitable bayesan approach to the decision are derived following a more consistent theory.
Bibliographic reference. Federico, A. / Paoloni, Andrea (1995): "Parametric speaker recognition over large population of telephonic voices", In EUROSPEECH-1995, 329-332.