Data-driven extraction of intonation contour classes

Abstract

In this paper we introduce the first steps towards a new datadriven method for extraction of intonation events that does not require any prerequisite prosodic labelling. Provided with data segmented on the syllable constituent level it derives local and global contour classes by stylisation and subsequent clustering of the stylisation parameter vectors. Local contour classes correspond to pitch movements connected to one or several syllables and determine the local f0 shape. Global classes are connected to intonation phrases and determine the f0 register. Local classes initially are derived for syllabic segments, which are then concatenated incrementally by means of statistical language modelling of co-occurrence patterns. Due to its generality the method is in principal language independent and potentially capable to deal also with other aspects of prosody than intonation.