Due to data sparsity, discriminative training has not been successfully applied in the system of vocal password by now. This paper proposed a novel vocal password framework based on a specific feature processing strategy, in which the new feature was used to represent the distance measure and the problem caused by data sparsity could be solved to some extent. As a result, the vocal password was actually transferred from verification to binary classification and the discriminative training of these two models was successfully accomplished on the minimum classification error criteria. After fusing this discriminative system with the GMM-UBM system, the EER performance decreased to 4.48%, relatively 17.95% and 59.68% lower than the GMM-UBM and the DTW system respectively on the corpus including 60 speakers. This result shows us the new application of discriminative training in the vocal password system is feasible and effective.