In this paper, we describe a discriminative technique to determine an optimal HMM topology for the each of the models in a continuous speech recognition system such that the word error rate (WER) is minimized. In conventional model selection techniques such as Bayesian information criterion (BIC), the model complexity is determined without considering the other classes in a system. In our work, an optimal model topology is selected by considering how well a given model can discriminate examples of other classes from its own. By doing so, the estimated model parameters indirectly make sure that class separability is increased. In an earlier work [1], we have proposed this technique and experiments were carried out on an E-set. Presently, we extend it for building a syllable-based continuous speech recognition system. Preliminary experiments carried out on the TIMIT corpus show that a considerable reduction in WER can be achieved using the proposed technique over the BIC-based technique for model selection.