Piet Mertens
Centre for Computational Linguistics,
K.U.Leuven,
PB 33,
B-3000 Leuven,
Belgium
Piet.Mertens@arts.kuleuven.ac.be

Specifying intonation for text-to-speech synthesis of French

This paper presents an algorithm for the generation of intonation contours in synthetic speech. The intonation contour is specified as a sequence of tone units associated with the syllable chain. (An additional step is required to convert this symbolic representation into a sequence of pitch targets (fundamental frequency values) associated with the sounds in the utterance.)
Most current systems for prosody generation use a variant of the "chinks and chunks" approach, which is merely a grouping of content words and function words. The algorithm presented in this talk derives prosodic structure from the syntactic structure (the parse tree) of the sentence. This process involves several steps in order to take into account the morphological, syntactic, rhytmic and phonetic properties and/or relations of the utterance.
The processing steps aim at:
- the identification of particular syntactic constructions (e.g. cleft);
- stress group formation;
- grapheme-to-phoneme conversion and syllabification;
- intonation group formation: accent groups are merged on the basis of syntactic relation and syllable count;
- reorganisation of this syntactico-prosodic structure taking into account rhythm and hierarchical limitations.