User Customized HMM/ANN Based Speaker Verification

In this paper, we describe a new speaker verification approach, using a hybrid HMM/ANN system, and accommodating user customized passwords. This system is exploiting the high phonetic recognition rates usually achieved by HMM/ANN speaker independent systems to infer the HMM topology associated with the user specific password from a few utterances of that password. Different adaptation schemes are then compared to quickly adapt the speaker independent ANN parameters used for HMM inference into speaker dependent parameters used for speaker verification. Different scoring criteria, based on normalized accumulated posterior probabilities (previously used as confidence measures in speech recognition) are also compared. Based on these improvements, our best system achieved false acceptance and false rejection rates of 8.2% and 3.2%, respectively, corresponding to an a posteriori threshold set to the minimum of the HTER (half total error rate), and in the worse case where all customers are using the same password.