Automatic prosody labeling of read norwegian

In this paper we present initial work on a method for automatic stress and boundary labelling of read EastNorwegian. The context of this work is automatic corpus annotation for unit selection speech synthesis. A phonological model of Norwegian prosody is described. The identification of syllable stress and major intonational boundaries are key prosodic events for building a prosodic description of a Norwegian utterance according to this model. A CART based method for automatic classification of syllable stress is presented. Initial experiments show that the method is capable of classifying syllables as unaccented or accented with high accuracy. 92.1 % of the unaccented syllables and 90.4 % of the accented syllables were correctly classified.