Adaptation robuste de modeles HMM pour la verification du locuteur dependante du texte

When deploying a secure system based on speaker verification, the limited amount of training data is usually critical. Indeed, the enrollment procedure must be fast and user-friendly. An incremental training of HMM speaker models, based on a MAP (Maximum A Posteriori) adaptation technique is used in order to make the enrollment more robust with only one or two utterances of the client password. This paper presents the improvements which can be achieved, in term of verification performance and stability of the decision thresholds. Our results highlight the benefits of MAP adaptation in conjunction with a synchronous alignment approach.