Investigation of lexical f0 and duration patterns in French using large broadcast news speech corpora


Rena Nemoto, LIMSI-CNRS and Université Paris-Sud 11
Martine Adda-Decker, LIMSI-CNRS
Jacques Durand, CLLE-ERSS/CNRS and Université de Toulouse-Le Mirail

This work aims at improving our knowledge of links between prosody and pronunciation variants in French. An original methodology is proposed to study prosodic regularities of French words via average f0 profiles, by making use of automatic processing and 13 hours of broadcast news speech. Investigated influential factors include word syllable length, duration, word-final schwa, parts of speech. The following questions are addressed: can specific lexical f0 profiles be measured automatically using large corpora? If so, how do they vary with respect to the cited influential factors? Results confirm the known tendency of word-final syllable accentuation. They also highlight some word-initial accentuation. Higher average f0 profiles are measured for increasing segment durations (locally decreasing speaking rate), but also for words ending with schwas. Future studies include phrase boundary annotation and the extension to a larger variety of speaking styles and languages.