Prosodic phrasing with inductive learning

Prosodic phrasing is an important component in modern TTS systems, which inserts natural and reasonable breaks into long utterance. This paper reports the study of applying several inductive machine-learning algorithms to prosodic phrasing in unrestricted Chinese texts. Two feature sets are carefully selected considering the effectiveness and reliability of them in practice. Then features and target boundary labels are extracted from a prepared speech corpus and used as training examples for inductive learning algorithms such as decision tree (C4.5), memory-based learning (MBL) and support vector machines (SVMs). The paper places emphasis on the comparison of the performance and speed of different learning techniques by training and testing them on the same corpus. The experiments show that all the algorithms achieve comparable results for both prosodic word and phrase prediction. It seems that prosodic word can be predicted from Chinese texts more accurately than prosodic phrase when using the same features and learning technique. Inductive learning is a promising way to prosodic phrasing, but it’s more important to find out good features than to apply different learning algorithms in order to improve the prediction accuracy dramatically.