Temporal control and training selection for HMM-based system

Most speaker-independent acoustic-phonetic decoding systems are based on hidden Markov models. Such systems lack a real temporal control for the phonetic models. Furthermore, inter-speaker variability makes speaker adaptation necessary. In order to solve these problems, we introduce two original approaches. On the one hand, discontinuities detected with the ForwardBackward Divergence method are used to constrain phonetic transitions and to perform a more accurate temporal control. On the other hand, an efficient interspeaker measure, based on AR-vector models, allows the selection of a speaker neighbourhood and the adaptation of the phonetic models. The contribution of these two methods is estimated on the TIMIT database.