Hidden Markov models are widely used for recognition algorithms (speech, writing, gesture, ...). In this paper, a classical set of models is considered: state space of hid- den variable is discrete and observation probabilities are modeled as Gaussian distributions. The models parame- ters are generally estimated with training sequences and the Baum-Welch algorithm, i.e. an expectation maxi- mization algorithm. However this kind of algorithm is well known to be sensitive to its initialization point. The problem of this initialization point choice is addressed in this paper: a model with a very large number of states which describe training sequences with accuracy is first constructed. The number of states is then reduced using a k-mean algorithm on the state. This algorithm is com- pared to other methods based on a k-mean algorithm on the data with numerical simulations.