 |
ESCA Workshop on ProsodyLund, Sweden |
 |
|
ESCA Workshop on ProsodyLund, Sweden |
|
We propose here a methodology and tools for the semi-automatic constitution of a intonative generation module..
The first stage of the work was the analysis of a corpus recorded by a reference speaker and based on a set of linguistic presuppositions. These presuppositions are based on the concept of some structural rendez-vous between the different levels of text on one part and the prosody, on the other part The processing of the data corpus was organized in a top-down hierarchy: sentences, clauses, groups and lexical units. The minimal symbolic unit is the syllable. For every level in the hierarchy, several initial classes ofFo contours are defined, each initially described by a maximal set of linguistic parameters. The validity of each class is first verified. Then the unification of the classes is systematically tested, using minimal pairs oppositions on the linguistic parameters. For every final class an average-contour is computed, which is a global form for this class. The result is a hierarchically structured dynamic contour lexicon of global intonative forms for which every representative is associated with a minimal set of distinctive attributes.
Generation of prosody then consists in the calculation of prosodic patterns by the top-down cumulative superpositions of contours taken from the lexicon. An application is the automatic generation of prosody in a text-to-speech synthesis system, which must be adapted to a given application.
Bibliographic reference. Aubergé, Véronique (1993): "Prosody modeling with a dynamic lexicon of intonative forms: application for text-to-speech synthesis", In Prosody-1993, 62-65.