Accuracy Improving Methods for Parametric Trajectory Modeling and Its Use in A* Search

In this paper, we first address the measurements to improve classification accuracy for parametric trajectory modeling (PTM), exploring the effect of context -dependent information, prosody knowledge (pitch, duration) and derivative features (to depict speech dynamics further besides the advantage of PTM on this aspect). Experiment shows 61.585% error reduction with these techniques. We then use it in the A* search for continuous Mandarin digit recognition. Here two implementations are introduced. First PTM score is linearly combined with HMM acoustic score as the cost function of partial path. This method did not influence result much. Then PTM score is used as confidence measure in A* search. After PTM validating, if the HMM result has a high confidence, a high weight is given for the HMM score in the cost function and Vice Versa. This time we get 7.81% string error reduction for uniphone model and 12.15% for triphone model, and the corresponding del-error, in-error and sub-error degrade too.