Prosody phrase break prediction in Vietnamese using decision tree

The intelligibility of synthesized speech is satisfactory, but the naturalness is fair in Vietnamese speech synthesis system without prosody phrase breaks. In order to improve the naturalness of synthesized speech, prosody phrase (L3) breaks are automatically predicted by using C4.5 decision tree algorithm in this paper. Firstly, we collect Vietnamese text and construct corpus. Then we obtain training date and testing data after word segmentation, part of speech (POS) tags and manual label of L3 breaks for the sentences in the corpus. Word segmentation and part of speech (POS) tags are conducted by applying text analysis software. Secondly, we extract the relevant attribute from the training data, and then obtain decision tree by using C4.5 decision tree algorithm. According to the pruned decision tree, L3 breaks are predicted in prosody labeling stage. Finally, we conduct objective and subjective test to the prediction. The results of evaluation show that an F-Score of 59.96% and acceptable rate of 70.6% can be achieved for the L3 prediction in closed set, and there is an F-Score of 58.37% and acceptable rate of 68.9% in open set. This experiment for the further improvement of naturalness of synthesized Vietnamese speech lays a foundation.