Speaker verification using phoneme-based neural tree networks and phonetic weighting scoring method

A text-dependent speaker verification system based on neural tree network (NTN) phoneme model and phonetic weighting scoring method is presented. The system uses a set of concatenated NTNs trained on phonemes to model a password. In contrast to the conventional stochastic approaches which model the phonemes by hidden Markov models (HMMs), the new approach utilizes the discriminative training scheme to train a NTN for each phoneme. The phoneme-based NTN is trained to discriminate the phoneme spoken by the speaker with respect to those spoken by other speakers. A weighted scoring method depending on the phoneme's ability for speaker verification is used to improve the performance. The proposed system is evaluated by experiments on the YOHO database. Performance improvements are obtained over conventional techniques.