The mostly used rule, i think, is the likelihood. You plot likelihood of each model vs number of states and make a choice at some maximum. At least that was what i have applied..If you have the time and nerves to make it even more robust, make a cross validation test to each case-model and them plot.

Choose the number of states which maximize the likehood is the simple. At least I have applied. Please feel free to see my paper.

Increasing the number of states will lead to a higher likelihood.  Unfortunately, increasing the number of states will also lead to a huge increase in the number of parameters (transition probabilities, initial state probabilities). Therefore, you should use a criterium which takes the number of parameters into account, for instance AIC or BIC.

One way to deal with this problem is to use the BIC score which uses what most people have mentioned so far (likelihood as a measure of fit) however, it penalizes complexity (increased number of states)

http://en.wikipedia.org/wiki/Bayesian_information_criterion