Describing intonation with a parametric model

Abstract

In this study a data-based approach to intonation modeling is presented. The model incorporates knowledge from intonation theories like the expected types of F 0 movements and syllable anchoring. The knowledge is integrated into the model using an appropriate approximation function for F 0 parametrization. The F 0 parameters that result from the parametrization are predicted from a set of features using neural nets. The quality of the generated contours is assessed by means of numerical measures and perception tests. They show that the basic hypotheses about intonation description and modeling are in principle correct and that they have the potential to be successfully applied to speech synthesis. We argue for a clear interface with a linguistic description (using pitch-accent and boundary labels as input) and discourse structure (using pitch-range normalized F 0 parameters), even though current text-to-speech systems usually still do not have the capability to predict most of the appropriate information.