GoalGetter: predicting contrastive accent in data-to-speech generation

This paper addresses the problem of predicting contrastive accent in spoken language generation. The common strategy of accenting 'new' and deaccenting 'old' information is not sufficient to achieve correct accentuation: generation of contrastive accent is required as well. I will discuss a few approaches to the prediction of contrastive accent, and propose a practical solution which avoids the problems these approaches are faced with. These issues are discussed in the context of GoalGetter, a data-to-speech system which generates spoken reports of football matches on the basis of tabular information.