Automatic prosody labeling using both text and acoustic information

Abstract

Prosody is an important factor for a high quality text-to-speech (TTS) system. The prosody is often described with a hierarchical structure. So the generation of the hierarchical prosody structure is very important both in the corpus building and the run-time text analysis. But the prosody labeling procedure is laborious and time consuming. Moreover, to keep the consistence between different labelers and even the same labeler in different time is difficult. In this paper an automatic prosody labeling system is presented, in which the decision tree plus Viterbi decoding framework proposed Wightman and Ostendorf (1994) is used. In the system, not only the acoustic information but also the text information such as the part-of-speech (POS) of a word is used. A prosody model is built up using the automatically labeled corpus for our Mandarin TTS system. Listening test shows that the automatic prosody labeling system works pretty well.