Estimation rapide de modèles de Markov semi-continus discriminants
Georges Linares, Christophe Levy, Jean-Christophe Plaignol.

In this paper, we present a fast estimation rule for MMIE (Maximum Mutual Information Estimation) training of semi-continuous HMM (Hidden Markov Models). We first present the method proposed by cite{Pov99} for weigth re-estimation. Then, the weight updating rule is formulated in the specific framework of semi-continuous models. Finally, we propose an approximated updating function which requires very low computational ressources. The first experiment validates this method by comparing our fast MMI estimator (FMMIE) and original one. We observe that, on a task of digit recognition, FMMIE obtains similar results than whose obtained by using the full updating rule. Then, semi-continuous models are integrated in a Large Vocabulary Continuous Speech Recognition (LVCSR) system. We use the real-time engine which has been involved in the ESTER cite{Gra04} evaluation campaign. Results show that the proposed MMIE models outperform significantly the system based on semi-continuous models and MLE training, while reducting the model complexity.