Modeling speech adaptation to altered sensory feedback through continuous learning of internal sensory predictions

The paper presents a version of an optimization-based model of speech production that reproduces key acoustic and articulatory features of motorsensory adaptation to altered sensory feedback. In the presented approach, the mechanism of motorsensory adaptation is based on regular updates, based on the sensory feedback perceived by the speaker, of two of the speakers internal models used for computing (near)-optimal articulation. These internal models, modeled as separate Artificial Neural Networks, are 1) a model that predicts the acoustic consequences of motor (articulatory commands) and 2) a model that predicts the somatosensory sensations from given motor commands. The paper presents simulations of adaptation experiments that successfully reproduce key acoustic and articulatory features of motorsensory adaptation of speech to altered sensory feedback. These include gradual and incomplete motorsensory adaptation when the auditory (or the somatosensory) feedback is suddenly altered (F1-shifted for the altered auditory feedback, forced jaw movement for altered somatosensory feedback). The presented simulations also show that the rate and magnitude of adaptation behavior depend on a small number of parameters. Variation in the values of these parameters can potentially explain inter-speaker differences in terms of adaptation behavior, including sensory preference.