Prosody Generation in the SmartKom Project

This paper describes the intonation prediction algorithm used in the speech synthesis component of the SmartKom dialog system. Both prosodic phrasing and default accent prediction operate on syntactic structure generated by a language generation module. Two configurational rules are applied to find candidate prosodic phrases. A harmonization algorithm selects the best candidates taking into account rhythm as well as a parameter specifying the optimal length of prosodic phrases. Depth of embedding and location within the syntactic structure determine the default accentuation, which can be modified depending on semantic factors.