Bayesian bpproach based decision in speaker verification

Considering Bayesian decision framework applied in the context of speaker verification, this paper presents a new way of handling troublesome anti-speaker model by proposing a redefinition of hypotheses involved in the classical statistical hypothesis test. This new definition of hypotheses is then implemented through a speaker independent normalization technique, named MAP approach. Besides supporting these new hypotheses, MAP approach takes the advantages of projecting likelihood scores into a probabilistic domain and therefore of providing the decision threshold with bounded and meaningful values. In this paper, different variants of MAP approach are presented which mainly aims at reducing likelihood variability, well-known in speaker verification to degrade system performance. MAP approach is firstly combined with classical normalization techniques (likelihood ratio normalization (world model) and/or Hnorm normalization technique). The second kind of variants consists in redesigning MAP approach to become speaker dependent. Experiments conducted on a subset of Switchboard database involving these different variants have showed that MAP approach is able to perform as well as classical normalization techniques while yielding probabilistic scores suitable for the decision threshold setting or the fusion of recognizer scores in the context of a multi-recognizer architecture.