Prosodic Word Grouping in Mandarin TTS System

This paper reports the methodology and results of prosodic word grouping for a Mandarin TTS system developed by the Fujitsu Laboratories. In view of any inner prosodic word break will make speech unintelligible or unnatural, a new prosodic word grouping framework is proposed. The word segmentation result can be regarded as an initial prosodic word sequence with grids inserted into each word boundary. The target of prosodic word grouping is to remove those grids that are assessed needlessly in the prosodic word level. Several prosodic word grouping methods can work together because each makes its own decision. As one of three methods used in our TTS system, a binary prosodic tree method is also proposed in this paper. Lastly, experiment results and discussion are presented.