Discriminatory measures for speaker recognition

This paper investigates methods for incorporating discriminatory information into speaker recognition systems. In particular, this information is used to supplement non-discriminative modeling approaches, such as dynamic time warping and hidden Markov modeling. The discriminative information is obtained from the neural tree network (NTN) and is integrated with the non-discriminative models via data fusion. Here, the outputs of each model are combined with two data fusion methods know as the linear opinion pool and log opinion pool. These methods are evaluated for text dependent speaker verification for two databases. For both experiments, the consensus driven system outperformed the systems based on individual models