This paper presents the use of a simulated annealing technique during the parameters estimation of a Hidden Markov Model (HMM) in a speech recognition system. This technique allows to move out of a local optimum which characterizes a classical Expectation Maximization (EM) algorithm, and thus to achieve a better estimation with a limited amount of training data. We choose here the Simulated Annealing Expectation Maximization (SAEM) algorithm introducing a simulated annealing technique in the EM method. The SAEM algorithm is compared to the classical EM algorithm, for both task- independent and task-dependent Viterbi training. The evaluation leads to significant improvement of recognition performances.