Context-dependent modeling is a well-known approach to increase modeling accuracy in continuous speech recognition. The most common way to implement this approach is via triphone modeling. Nevertheless, the large number of such models results in several problems in model training, whilst the robust training of such models is often hardly obtained. One approach to solve this problem is via parameter tying. In this paper, clustering has been carried out on HMM state parameters and the states allocated to any cluster are tied to decrease the overall number of system parameters and achieve robust training. Two types of groupings, one based on the final trained model set parameters and their inter-model distances and the other based on the training data and a decision tree, have been carried out. In the implementation of the later, a decision tree based on the acoustic properties of the Persian (Farsi) language and the phonetic similarities and differences has been designed. The results obtained have shown the usefulness of both the approaches. However, the second approach has the advantage of making the estimation of unseen model parameters possible.