Predicting the Intonation of Discourse Segments from Examples in Dialogue Speech

Abstract

In the area of speech synthesis it is already possible to generate understandable speech with discourse neutral prosody for simple written texts. However, at ATR-ITL we are researching speech synthesis techniques for use in a speech translation environment. Dialogues, in such conversations, involve much richer forms of prosodic variation than are required for the reading of texts. For our translations to sound natural it is necessary for our synthesis system to offer a wide range of prosodic variability, which can be described at an appropriate level of abstraction. This paper describes a multi-level intonation system which generates a fundamental frequency (F ₀ ) contour based on input labelled with high level discourse information, including speech act type and focussing information, as well as part of speech and syntactic constituent structure. The system is rule driven but rules (and parameters) are derived from naturally spoken dialogues. Two experiments using this model are described, testing its accuracy. First results are given for a system to predict ToBI intonation labels from discourse information use a CART decision tree. Second a detailed investigation of the intonational variation of the word “okay” in different discourse contexts is presented.