Tongue motor control: deriving articulator trajectories and muscle activation patterns from an optimization principle

International audienceSpeech motor control is known to be highly resistant to perturbations of various origins. In particular several studies have demonstrated the capacity of speech motor control system to deal with steady state perturbations such as bite-block, lip-tube, or modified palatal vault, using the excess degrees of freedom of the relations between motor commands and articulatory positions or between articulatory positions and spectral characteristics of speech sounds. Efficient adaptations to repeated perturbations of the auditory-feedback, of the jaw movements, or of vocal tract morphology have also been shown. This flexibility echoes what is observed for upper limb motion and suggest the operation of a similar control principle.In this study, we explored the predictions of optimal feedback control theory applied to tongue motion. We assumed that the CNS aims to minimize both neuromotor effort and motor error. We also assumed that it estimates the current state of the articulator system through the combination of delayed and noisy sensory feedbacks and the predictions of an internal model of the dynamics of the speech production system. In more detail, we used a two-dimensional finite element model of the tongue, or in some simulations a reduced model of it. The first three formants of the voice were calculated from the vocal-tract shape with a harmonic model of the vocal tract. For simplicity, functions of tongue position and velocity were considered as proprioceptive input. We assumed that motor commands and sensory signals were corrupted by multiplicative noise. Feedback updated the state of the system through an extended Kalman filter. We essentially explored the generation of movements from a rest (e.g. schwa-like) tongue position to vowels like [i], [a] or [ͻ], or between wovels. Optimization predicted complete muscular activation and kinematic patterns in time. Movements of flesh points were roughly sigmoidal, in accordance with reported articulatory recordings. Muscle activation was consistent with the findings of tongue EMG studies; tongue being a muscular hydrostat and not an articulated rigid body, muscle dynamics were specific: most movements involved a long initial bell-shaped activation of agonists followed by a final short braking antagonist burst. Interestingly, trajectories in acoustic space were sometimes markedly curved (esp. in F1-F2 space). Variability in final acoustic production emerged as a consequence of sensory and motor noise, for a fixed goal and starting posture, providing an estimate of the sensorimotor contribution to phonemic variability. Specific experimental work using these artificial speech task will have to be undertaken to validate the kinematic and acoustic predictions of this model