Score normalization technique for text-prompted speaker verification with Chinese digits

Abstract

A text prompted speaker verification system is presented in this paper. This system is based on ten Chinese digits. Basic acoustic models are speaker dependent and content dependent phoneme HMMs which were generated by adapting speaker independent models to the utterances of specific speakers. An obvious constraint for normalization techniques used in TDSV is that the phrases with the same content should be used for competitive cohort models. So many of the score normalization techniques are either difficult to implement because of lack of data or not good for performance improvement because of poor estimation of the normalization parameters. We propose a method which combines the traditional T-Norm and Cohort Norm together to find a good tradeoff of testing utterance normalization and target speaker model normalization. The proposed method improved the system performance from the baseline equal error rate 3.42% for T-Norm and 2.72% for Cohort Norm to 2.50%. © Springer-Verlag Berlin Heidelberg 2007.