Coarticulation in a whole event model of speech production

Previous models of coarticulation have used varying combinations of advance planning and on-line calculation of weighted averages to determine how temporally overlapping speech sounds interact [see Farnetani and Recasens, Hbk. Phonet. Sci. Wiley-Blackwell (2010)]. A robust model of coarticulation should be able to predict such local interactions, as well as to describe changes resulting from rate variation, which should influence degree of temporal overlap between adjacent events. Using a model of tongue-jaw-hyoid biomechanics [Stavness et al., J. Biomech. (2012)], the present paper demonstrates that typical cases of lingual coarticulation can be attributed to the intrinsic biomechanics of the human body in an entirely feed-forward model with no additional machinery. Biomechanical modeling outcomes are compared to speech articulations at different speech rates, and show that naturalistic coarticulatory patterns emerge simply by varying degrees of temporal overlap in a biomechanically realistic model. The built-in mechanics of the human body can handle coarticulatory interactions with no extrinsic model at all, save one that identifies a) the right body parts, and b) the time-course of events. Results are interpreted in a “Whole Event” model of speech. [Research funded by NSERC.]