Acoustical segmental duration or articulatory inter-targets as an indicator of speaker specific kinematic properties

摘要

The segmental duration is an easily measurable speech parameter on the acoustic signal. Recent studies have shown that segmental duration is speaker specific (Pfitzinger, 2002), and it can be used for the automatic speaker recognition exploiting this speaker specificity (Ferrer et al., 2003). In this report, we discuss the interest in the temporal aspects of speech production in the context of the acoustic-to-articulatory inverse. In fact, its characteristic of speaker specificity suggests its possible link with the kinematic and underlying bio-mechanical properties specific to individual speakers. Everything else being equal, a longer segmental duration can be regarded as the manifestation of either a longer path length between two successive articulatory targets or a slower articulator's speed suggesting a weaker stiffness of the related muscles in bio-mechanical terms. Turning our attention to the inverse problem, the derived kinematic properties may allow us to adapt the control sequence to a specific speaker in connection with a generic articulatory model already adapted to the morphology of that speaker. Moreover, the acoustically derived bio-mechanic properties can provide a reasonable constraint on the possible articulatory trajectories in the speech inversion. In this sturdy, we shall focus our attention to unvoiced sibilant fricatives, /s/ and /∫/, because their segmental duration can be automatically and reliably measured on a large speech database. Actually we have formulated a robust segmentation method with high accuracy.