Sesile Metin Bagimli Kimlik Dogrulama

In this paper, a text dependent speaker verification system has been proposed. The utterances have been aligned and segmented via the HMM of the utterance of the claimed speaker which is constructed based on a speaker independent phoneme model to be verified. Then the distance between the feature vector of this aligned utterance and the mean utterance vector due to the reference utterances of the claimed speaker have been compared for verification. The equal error rate is observed to be under 7% for this system where 600 utterances from 30 people have been processed. Using a text dependent system and user specific threshold values for verification are two factors that have increased success rates.