Prosody-based sentence boundary detection in Chinese broadcast news

In this paper, we explore the use of prosodic features in sentence boundary detection in Chinese broadcast news. The prosodic features include speaker turn, music, pause duration, pitch, energy and speaking rate. Specifically, considering the Chinese tonal effects in pitch trajectory, we propose to use tone-normalized pitch features. Experiments using decision trees demonstrate that the tone-normalized pitch features show superior performance in sentence boundary detection in Chinese broadcast news. Furthermore, feature combination is able to achieve apparent performance improvement by intuitive feature interactive rules formed in the decision tree. Pause duration and a tone-normalized pitch feature contribute the most part of the feature usage in the best-performing decision tree.