Text-dependent Speaker Verification Using Word-based Scoring

As the separated modeling methods are widely used in text-dependent speaker verification task. The reason why they are so effective is discussed in this paper. A word-based scoring method is then proposed based on our discussion. Specifically, a segmentation algorithm is firstly used for segmenting the enrollment and test utterances into words, automatically. Then every segment of the enrollment utterance is used to enroll a word based speaker model. Scoring is done with each testing segment and a corresponding word-based speaker model of the same word. The experiments are carried out on a short duration text-dependent speaker verification database in Chinese spoken language. Our examples show that, systems based on word-based scoring method are superior to the relevance MAP GMM-UBM system and achieve significant performance improvement.