Several attempts are made to improve the performance of a handwritten cursive English recognizer which has been developed by interconnecting hidden Markov models (HMMs). First, we have shown that a structural constraints should be observed regarding self-loop at termination nodes in order to interpret the HMM output as a probability measure. Second, two measures are proposed to evaluate the degree of training in terms of specialization and generalization. The measures are utilized to train individual HMMs better by getting insights about the proper size of training set. The last attempt is on the better selection of character models. Instead of creating one model for each alphabet, data clustering algorithm is used to determine proper number of models and data grouping. Experimental results show that the proposed methods are useful to improve recognition accuracy.