International Symposium on Chinese Spoken Language Processing (ISCSLP 2002)

Taipei, Taiwan
August 23-24, 2002

Decision Tree Based Unit Pre-Selection in Mandarin Chinese Synthesis

Zhen-Hua Ling, Yu Hu, Zhi-Wei Shuang, Ren-Hua Wang

iFly Speech Laboratory, University of Science and Technology, Hefei, China

In this paper we introduce a classification and regression tree (CART) based method to improve the efficiency of our corpus-based Mandarin Chinese synthesis system and at the same time maintain the quality of the synthesized speech. CART is one kind of the popular decision tree, through which, the candidates of the same tonal syllable in the corpus are pre-classified in three different ways, taking into account their experiential rule distance, segmental features or prosodic features separately. The difference of these methods exists in the different measurement of the distance between any two candidates, while the distance is used to construct the decision tree. The implementation and comparison of these three kinds of unit pre-selection methods mentioned above and their results are presented. And finally we come to the conclusion that prosodic characteristics of syllables are more important than segmental characteristics in Mandarin Chinese synthesis.

Full Paper

Bibliographic reference.  Ling, Zhen-Hua / Hu, Yu / Shuang, Zhi-Wei / Wang, Ren-Hua (2002): "Decision tree based unit pre-selection in Mandarin Chinese synthesis", In ISCSLP 2002, paper 58.