Text-independent speaker verification with a multiple binary classifier model

Describes the adaptation of a text-dependent talker verification approach to a text-independent system. In the adapted model, an (N-1) set of Moody-Darken radial basis function networks is trained for each of N true talkers from a reference database. Each network within a set is trained with the corresponding true talker´s parametrised speech and a single alternative talker is chosen randomly. When an identity is claimed, the (N-1) networks are all tested with the claimant´s utterance. Acceptance rates are averaged over the set. Should this resulting mean exceed a true talker-dependent threshold, the unknown talker is accepted. Otherwise, the talker is rejected as an impostor. Results obtained from 100 imposture attempts show that the system has the potential of being foolproof, under ideal operating conditions