This paper presents a work-in-progress on the automatic analysis of discourse genre in non-elicited speech. The study is focused on the development of bottom-up methods for automatic validation of discourse typologies found in linguistic descriptions (prosodic, syntactic, pragmatic and/or contextual and situational cues). The linguistic classification examined here opposes five discourse genres +/- controlled. To test this a priori classification under prosodic criteria, we propose a method that provides an automatic and dynamic estimation of discourse genre typology i.e. of prosodic similarities between discourse genres. This is achieved in a two-step procedure: a set of discriminant prosodic patterns is estimated and then used to raise a typology of discourse genres based on prosodic similarity criterion. The discriminant analysis reveals that a small number of prosodic patterns is sufficient to discriminate the 5 discourse genres. The typological analysis reveals some multi-level caterogical oppositions on a continuous prosodic scale that can be interpreted in terms of +/- controlled speech.
Bibliographic reference. Obin, Nicolas / Lacheret-Dujour, Anne / Veaux, Christophe / Rodet, Xavier / Simon, Anne-Catherine (2008): "A method for automatic and dynamic estimation of discourse genre typology with prosodic features", In INTERSPEECH-2008, 1204-1207.