The phonetic line of the text-to-speech synthesis system we are working on, is classically generated by a rules/lexicon based grapheme-to-phoneme conversion, and then by the concatenation of polyphones (including some consonant clusters).

We present in this paper the prosodic line of this system. The basic idea is that each specific application of the synthesis requires a particular prosodic strategy. A complete system was elaborated to handle at the same time the textual data and the prosodic data of the application specific corpus. The characteristics of such a system are the automatic extraction, visualization, and averaging of some prosodic contours on the basis of prosodic and/or textual criteria. The results are a lexicon of mean prosodic contours, only parametrized by mean linguistic events.