Can We Tell Apart Intonation From Prosody (if We Look At Accents And Boundaries)

Studies on prosody/intonation normally look for important (distinctive) features denoting linguistic contrasts in production or perception experiments. The recent development in automatic speech processing and the availability of large speech data bases made it possible to have a fresh look at these topics. In our study, we classify automatically accent and boundary positions in a spontaneous speech corpus with a large feature vector comprising as many relevant prosodic features as possible. The results obtained for diierent subsets of prosodic features (F0, duration, energy, etc.) show that each feature class contributes to the marking of accents and boundaries, and that the best results can be achieved by simply using all feature subsets together. Finally, we discuss possible conclusions for prosodic theory and for the application of prosody in speech processing.