The “kiel Corpus of Read Speech” as a Resource for Prosody Prediction in Speech Synthesis

The naturalness of synthetic speech depends strongly on the prediction of appropriate prosody. For the present study the original annotation of the German speech database “Kiel Corpus of Read Speech” was extended automatically with syntactic features, word frequency, and syllable boundaries. Several classification and regression trees for predicting symbolic prosody features, postlexical phonological processes, duration, and F0 were trained on this database. The perceptual evaluation showed that the overall perceptual quality of the German text-to-speech system MARY can be significantly improved by training all models that contribute to prosody prediction on the same database. Furthermore, it showed that the error introduced by symbolic prosody prediction perceptually equals the error produced by a direct method that does not exploit any symbolic prosody features.