A new method of building decision tree based on target information

The algorithm of unit pre-selection based decision tree has been adopted in the corpus-based TTS system, but the effect of the decision tree, which is built by conventional methods, is not satisfactory. This paper proposed a new method of building a decision tree based on a mass of target information obtained in the synthesizing course. By synthesizing a large quantity of text and logging the information of target units, we used these target information as training data to classify the corpus unit and build the decision tree. The splitting criterion is to ensure that the best units of target units correctly classified, which guarantee that the synthesis quality would not be reduced. And the decision tree built by this method has an advantage that it has equal units among the different classes. By evaluating the synthesis quality in objective and subjective measure, this method is better than the conventional ones.