Automatic labeling of large prosodic databases: Tools and methodology

Construction of synthetic prosodic models for high‐quality speech synthesis implies the association of linguistic information and prosodic information on a corpus especially designed to study certain linguistic insights [Emerard and Benoit, 16emes Journees d'Etudes sur la Parole, Hammamet, Tunisie, 1987, pp. 224–226]. This paper concerns the automatic acquisition of quantitative description of the prosodic contours compatible with a high‐quality synthesis approach. Originality of this approach is that the sound continuum's description is based on a phonetic model, i.e., phonemes as emergence functions issued by a temporal decomposition (TD) technique [Bailly, Marteau, and Abry, Proc. Int. Conf. ASSP, Glasgow, Scotland, 1989, pp. 508–511]. The speaker‐independent automatic alignment [Wang et al., J. Acoust. Soc. Am. Suppl. 1 87, S106 (1990)] and F0 tracker are presented: They deliver a prosodic contour description (duration, nucleus duration, F0 contour, energy). This description is sufficient to describe ...