An optimality-guided model of differential adaptation to opposing perturbations

This paper introduces an artificial neural-network-based computational model of differential sensorimotor adaptation to opposing auditory feedback perturbations in speech. A neural Optimality Predictor generating (near-)optimal articulation plans is trained to balance maximal intelligibility and minimal articulatory effort, based on lexical, phonemic, and stylistic context. Adaptation is driven by two main mechanisms: updates of internal sensory predictors estimating sensory outcomes of articulatory plans, and updates of the Optimality Predictor. Simulations replicate experimentally observed differential adaptation, predicting context-dependent shifts in articulatory and acoustic outputs. The model demonstrates that target lexical goals cluster by phonological representations, with vowels distributed in the model's embedding space according to their articulatory and acoustic properties. During adaptation, shifts in this space correspond to changes in produced vowel features. The model explains 61% of the variability in adaptation through embedding distance between to-be-perturbed words, accounting for reduced adaptation in homophones. Overall, our findings suggest that differential adaptation is primarily driven by differences in the cognitive representations of target productions, whether differences are lexical, phonological, and/or phonetic in nature.