Simplification of mixture models has recently emerged as an important issue in the field of statistical learning. The heavy computational demands of using large order models drove researches to investigate how to efficiently reduce the number of components in mixture models. The simplification, in solutions proposed so far, was performed by maximizing a certain measure of similarity to the original model, regardless of the discriminative qualities among models of different classes. This paper proposes a novel discriminative learning algorithm for reducing the order of a set of mixture models. The suggested algorithm is based on maximizing the correct component association. Experiments, performed on acoustic modeling in a basic phone recognition task, indicate that the proposed algorithm outperforms the comparable non-discriminative simplification algorithm.