A bootstrapping approach to automating prosodic annotation for limited-domain synthesis

Most speech synthesis systems use symbolic prosody labels for marking emphasis and phrase structure. but in corpus-based approaches prosodic annotation of speech is a labor intensive process driving up the cost of development of new voices. This paper explores the potential for reducing that cost by using a bootstrapping approach to automatic prosodic annotation, particularly in a limited domain application. A perceptual experiment shows that using predominantly automatic prosody labels we can achieve nearly as high synthesis quality as if all data was hand-labeled.