Hidden Markov Models (HMMs) are an useful and widely utilized approach to the modeling of data sequences. One of the problems related to this technique is finding the optimal structure of the model, namely, its number of states. Although a lot of work has been carried out in the context of the model selection, few work address this specific problem, and heuristics rules are often used to define the model depending on the tackled application. In this paper, instead, we use the notion of probabilistic bisimulation to automatically and efficiently determine the minimal structure of HMM. Bisimulation allows to merge HMM states in order to obtain a minimal set that do not significantly affect model performances. The approach has been tested on DNA sequence modeling and 2D shape classification. Results are presented in function of reduction rates, classification performances, and noise sensitivity.