Automatic analysis of prosodic structure statements of various styles

Abstract

The aim of this study is to characterize automatically the boundary tones in French, from the point of view of their categories, their acoustic realizations and their distribution according to a prosodic grammar. If the automatic labelling proves pertinent, the proposed method would lead to an automatic process to determine the prosodic structure of speech recordings. From the C-PROM corpus, four spontaneous recordings with various styles (oral reading, university lecture, narration, political discourse) were automatically analyzed based on 1) the identification of prominent syllables (from the information already encoded in the corpus transcription), and 2) on the melodic contours glissando values relatives to prominent syllables. From na existing definition of prosodic contours in French, C0, C1, C2 and Cn, the validity of a prosodic gramar has been established, as reflected by the results shown in Table 1, as only one occurrence (*C2C0) not predicted by the grammar has been observed.