Détection automatique de frontières prosodiques dans la parole spontanée
Katarina Bartkova, Natalia Segal.

The present study addresses the issue of the automatic detection of prosodic units in French. An analysis of two prosodic parameters, phone duration and F0 slope values, carried out on two spontaneous speech data bases recorded by several thousands of speakers, revealed relevant deviations of these parameters on prosodic junctions. Vowel durations and F0 slopes are used to automatically detect prosodic units on the two data bases. The phone duration is modelled as the ratio of two subsequent vowel durations. Apart from the duration ratio, duration values are also modelled. The detection of prosodic units based on the F0 uses the value of the F0 slope and its standard deviation, recalculated after each pause. An evaluation of the automatic detection is carried out by comparing the prosodic border locations with the lexical boundaries and also with prosodic boundaries obtained by manual segmentation.