An automated process for building a prosodic structure form transcribed speech recordings in French is presented, based on the incremental prosodic model [1, 2, 3]. In this model, the prosodic structure is defined incrementally by dependency relations instantiated by melodic contours located on the last syllable of the last word of stress groups, subject to a rhythmic constrain limiting the gap between successive stressed syllable to a 250-1250 ms range. Although they frequently contain lexical words (noun, verb, adverb, adjective), stress groups in French can also include only grammatical words (pronoun, conjunction, preposition). Melodic contours are phonologically defined from their melodic rise or fall and their glissando value ensuring their function as dependency markers between stress groups. The algorithm proceeds from an orthographic transcription as follows: 1. Automatic segmentation of the orthographic text into IPA and word tiers 2. Automatic annotation of stressed vowels in three classes (followed by 250 ms silence, above the glissando threshold and lexical category based) 3. Assignment of melodic contours from fundamental frequency values at stressed vowels boundaries. Comparisons with automatic and manual stressed syllable annotation on existing corpora are given, showing the validity of the phonological rules implemented in the algorithm.
Prosodic structure in language understanding: Evidence from tone sandhi in Mandarin
