¹Key Laboratory of Trustworthy Computing, Shanghai, East China Normal University
²College of Computer and Information Engineering, Tianjin Normal University, Tianjin, China

---

Abstract—For a Chinese speech synthesis system, hierarchical prosody structure generation is a key component. The prosodic word, which is the basic prosodic unit, plays an important role in the naturalness and intelligibility of Chinese Text-To-Speech system. However, obtaining human annotations of prosodic words to train a supervised system can be a laborious and costly effort. To overcome this, we explore active learning techniques with the goal to reduce the amount of human-annotated data needed to attain a given level of performance. In this paper Active Maximum Entropy Markov Model(AMEMM) is used to predict Chinese prosodic word boundaries in unrestricted Chinese text. Experiments show that for most of the cases considered, active selection strategies for labeling prosodic word boundaries are as good as or exceed the performance of random data selection.

Index Terms—Prosodic Word, Text-to-Speech System (TTS), Active Learning, Maximum Entropy Markov Model.

[PDF]

Cite: Ziping Zhao, Xirong Ma, "Active Learning for Prediction of Prosodic Word Boundaries in Chinese TTS Using Maximum Entropy Markov Model," Journal of Software vol. 8, no. 12, pp. 3222-3228, 2013.