Evaluating the pronunciation of proper names by four French grapheme-to-phoneme converters

Philippe Boula de Mareüil (1) , Christophe d'Alessandro (1) , Gérard Bailly (2) , Frédéric Bechet (3) , Marie-Neige Garcia (4) , Michel Morel (5) , Romain Prudon (4) , Jean Véronis (6)

1 LIMSI - Laboratoire d'Informatique pour la Mécanique et les Sciences de l'Ingénieur
2 ICP - Institut de la communication parlée
3 LIA - Laboratoire Informatique d'Avignon
4 ELDA - Evaluations and Language resources Distribution Agency
5 CRISCO - Centre de recherche inter-langues sur la signification en contexte
6 DELIC - DEscription Linguistique Informatisée sur Corpus

This article reports on the results of a cooperative evaluation of grapheme-to-phoneme (GP) conversion for proper names in French. This work was carried out within the framework of a general evaluation campaign of various speech and language processing devices, including text-to-speech synthesis. The corpus and the methodology are described. The results of 4 systems are analysed: with 12-20% word error rates on a list of 8,000 proper names, they give a fairly accurate picture of the progress achieved, the state-of-the-art and the problems still to be solved, in the domain of GP conversion in French. In addition, the resources and collected data will be made available to the scientific and industrial community, in order to be re-used in future bench-marks.

Mots clés: Proper names; Evaluation; Text-to-speech; Pronunciation

Domaines: Linguistique

Philippe Boula de Mareüil, Christophe d'Alessandro, Gérard Bailly, Frédéric Bechet, Marie-Neige Garcia, et al.. Evaluating the pronunciation of proper names by four French grapheme-to-phoneme converters. Interspeech'2005 - Eurospeech. 9th European Conference on Speech Communication and Technology, Sep 2005, Lisbonne, Portugal. pp.1521-1524. ⟨hal-00103607⟩