Learning Intonation Rules for Concept to Speech Generation

Abstract

In this paper, we report on an effort to provide a general-purpose spoken language generation tool for Concept-to-Speech (CTS) applications by extending a widely used text generation package, FUF/SURGE, with an intonation generation component. As a first step, we applied machine learning and statistical models to learn intonation rules based on the semantic and syntactic information typically represented in FUF/SURGE at the sentence level. The results of this study are a set of intonation rules learned automatically which can be directly implemented in our intonation generation component. Through 5-fold cross-validation, we show that the learned rules achieve around 90% accuracy for break index, boundary tone and phrase accent and 80% accuracy for pitch accent. Our study is unique in its use of features produced by language generation to control intonation. The methodology adopted here can be employed directly when more discourse/pragmatic information is to be considered in the future.