This paper proposes an original approach to the task of speaker verification, in which the training process consists in a direct modeling of the score function. It divides the parameter space in disjoint regions where a score can be obtained as a simple function of the vector position in the region. The aim of this approach is, on the one hand to overcome some undesirable properties of the Gaussian Mixture Models (GMMs), and on the other hand, to speed up the decision process.
First, we present the formalism of probabilistic speaker verification and we discuss some motivations for exploring alternative approaches. We then describe a method currently under investigation, which is based on a binary recursive partition of the acoustic parameter space into regions to which an elementary scoring function is associated. Finally, we provide illustrations and preliminary results of the method, together with conclusions and perspectives.