AGMMA: A Novel Incremental Adaptation Method and its Application to Speaker Recognition

Classical adaptation approaches are generally used for model adaptation with a particular speaker or a specific environment. An incremental adaptation method is presented in this paper called AGMMA which is based on the modified segmental-EM algorithm and apply it to speaker recognition system. The initial model is trained on a limited amount of data and adapted recursively to enrich itself incrementally with the data available in each session. The proposed method evaluates the expectation of the initial data, which would be used in the segmental EM algorithm applied on both initial and new data, by the statistics of initial data. Experiments were taken on YOHO database that was a high quality microphone speech database and an attendance system that ran over eleven months. The results on YOHO database showed that AGMMA outperforms ARGMM and classical Bayesian adaptation in most of the cases. Significant profits are also achieved when AGMMA applied to the attendance system in real-life environment