Bayesian bpproach based decision in speaker verification.

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