Speech Motor Control combining the λ-model and Optimal Feedback Control - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Speech Motor Control combining the λ-model and Optimal Feedback Control

Résumé

The tongue is a muscular organ known to be crucial for speech production. Its control is complex because it behaves as a soft muscular hydrostat and its shape is significantly influenced by a relatively short delay somatosensory feedback (<50ms) comparable to a polysynaptic stretch reflex [Ito et al., 2023]. Thus, to model speech production, the GEPPETO model [Payan and Perrier, 1997] includes as the physical system to be controlled a biomechanical model of the tongue based on a deformable Finite Element mesh with seven muscles modeled, which generate active forces based on the λ-model [Feldman, 1986]. Thus, muscle force is controlled indirectly via astretch-reflex like mechanism that takes into account the difference between the current muscle length and a threshold length λ which is assumed to be the control variable specified by the Central Nervous System.In the original GEPPETO model, a phoneme sequence to be produced is specified in terms of phoneme-related target values of the λ commands and their timings. Importantly the GEPPETO model does not include any notion of intended or optimal tongue trajectory: the shaping over time of the tongue results from the interaction between the tongue biomechanical characteristics and the λ commands that vary at a constant rate of shift between their successive target values. This modeling approach has proved to realistically replicate physical properties of tongue movements in speech [Perrier et al., 2003, Perrier and Fuchs, 2008]. However, this model also has a significant limitation, since it is not possible to precisely control the accuracy with which an intended phoneme-related target equilibrium position is reached, spatially and in time. In contrast, the online integration of sensory feedback makes this accurate control possible in the stochastic feedback control theory ([Todorov and Jordan, 2002]), which is fully compatible with a discrete target-based sequence planning and the absence of trajectory optimization based on kinematic or force variables.Thereby, we implemented an Optimal Feedback Control (OFC) model to control the Tongue model, in which the λ commands are estimated via the minimization of a cost function that combines neuromuscular effort (i.e. λ changes over the sequence), and a penalty on accuracy in reaching phoneme-related goals defined in a multimodal (auditory and tactile) space. The OFC framework includes a prediction of the tongue shape and its associated sensory inputs thanks to an internal model of tongue biomechanics that is implemented as an LSTM network trained on more than 23000 simulations with the Finite Element model of the tongue. Consequently, trajectories reflect both a representation of biomechanics and a higher-order effort minimization.Results compare trajectories and muscle activations generated with the OFC and the GEPPETO models. Using OFC does not significantly affect the main characteristics of tongue trajectories and their velocity profiles. However, the control of rate and timing of speech movements has greatly improved. On top of that, anticipatory effects in /Consonant-Vowel-Consonant-Vowel/ sequences are nicely accounted for.
Fichier non déposé

Dates et versions

hal-04235981 , version 1 (10-10-2023)

Identifiants

  • HAL Id : hal-04235981 , version 1

Citer

Tsiky Rakotomalala, Pierre Baraduc, Pascal Perrier. Speech Motor Control combining the λ-model and Optimal Feedback Control. Progress in Motor Control XIV, Sep 2023, Rome (Italie), Italy. ⟨hal-04235981⟩
43 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More