Text chunking for intonational phrase prediction in Chinese

Two methods for intonational phrase (INP) prediction in Chinese TTS (text to speech) system are proposed, both of which exploit the information of base phrase (BP). Method 1 is called BP-based INP boundary prediction, in which, BPs in a sentence are recognized at first, then INP boundaries are predicted on the basis of words and recognized BPs. Method 2 is called BP-filtering INP boundary prediction, in which, recognized BPs are used to filter out some impossible boundaries. Decision tree (DT) was used as learning method for both BP recognition and INP boundary prediction. Comparing to the conventional method without BP recognition, the two methods proposed here achieved 3.6% and 5.6% reduction in the measure of unacceptability separately. More analysis shows that, by improving the performance of BP recognition, there is maximally 36.9% reduction in unacceptability.