An optimality-guided model of differential adaptation to opposing perturbationsa)

Area of Science:

• Computational Neuroscience
• Speech Science
• Artificial Intelligence

Background:

• Sensorimotor adaptation in speech is crucial for maintaining vocal communication.
• Understanding how the brain adjusts speech production in response to auditory feedback is a key challenge.
• Previous models have not fully captured the differential nature of adaptation across different linguistic contexts.

Purpose of the Study:

• To introduce a novel artificial neural network model for differential sensorimotor adaptation in speech.
• To investigate the mechanisms driving adaptation to opposing auditory feedback perturbations.
• To explore the role of cognitive representations in modulating speech adaptation.

Main Methods:

• Developed a computational model using artificial neural networks, including a neural Optimality Predictor.
• Trained the model to balance intelligibility and articulatory effort based on linguistic context.
• Simulated adaptation by updating internal sensory predictors and the Optimality Predictor.

Main Results:

• The model successfully replicated experimentally observed differential adaptation patterns.
• It predicted context-dependent shifts in articulatory and acoustic speech outputs.
• The model explained 61% of adaptation variability based on the cognitive representation (embedding distance) of words, including reduced adaptation for homophones.

Conclusions:

• Differential speech adaptation is primarily driven by cognitive representations of target productions (lexical, phonological, phonetic).
• The model highlights the interplay between sensory prediction updates and changes in the speech production planning.
• Findings suggest that the brain leverages abstract linguistic representations to guide flexible and adaptive speech motor control.