This paper presents a new discriminative approach for training Gaussian mixture models (GMMs) of hidden Markov models (HMMs) based acoustic model in a large vocabulary continuous speech recognition (LVCSR) system. This approach is featured by embedding a rival penalized competitive learning (RPCL) mechanism on the level of hidden Markov states. For every input, the correct identity state, called winner and obtained by the Viterbi force alignment, is enhanced to describe this input while its most competitive rival is penalized by de-learning, which makes GMMs-based states become more discriminative.Without the extensive computing burden required by typical discriminative learning methods for one-pass recognition of the training set, the new approach saves computing costs considerably. Experiments show that the proposed method has a good convergence with better performances than the classical maximum likelihood estimation (MLE) based method. Comparing with two conventional discriminative methods, the proposed method demonstrates improved generalization ability, especially when the test set is not well matched with the training set.