We publish first experiments on a new approach for training unigram prior probabilities of pronunciation variants using the Discriminative Model Combination (DMC) framework. A specific analysis of the approach is provided together with an error analysis before and after the DMC-training of the pronunciation priors.