Auditory, somatosensory, and motor interactions in speech acquisition and production

Frank H. Guenther

Dept of Cognitive & Neural Systems, Boston University, Boston, MA

This talk will describe the DIVA model of speech production and experiments designed to test and refine the model. The model is a neural network that learns to control movements of an articulatory synthesizer. Model components correspond to regions of the cerebral cortex and cerebellum that are active during simple speech production tasks. A babbling cycle is used to train neural mappings between phonological, articulatory, auditory, and somatosensory representations. These learned mappings encode speaker-specific information regarding the relationships between the different representations. After learning, the model is capable of producing combinations of the sounds it has learned by commanding appropriate movements of the speech articulators in the articulatory synthesizer. Computer simulations verify the model's ability to account for a wide range of experimental results concerning speech movements, including data on acquisition of speaking skills, perception-production interactions, coarticulation, and motor equivalence. The model also generates specific predictions that can be tested with electromagnetic articulometry and functional brain imaging.