The sensorimotor foundations of phonology: a computational model of early childhood articulatory and phonetic development

HABLAR is a computational model of the sensorimotor foundations of early childhood phonological development. It is intended to explain key characteristics of normal phonological development including the phonetic characteristics of babble, systematic and context sensitive patterns of sound substitutions and deletions, and overgeneralization of pronunciation patterns. HABLAR ("Hierarchical Articulatory Based Language Acquisition by Reinforcement learning") features an auditory system which categorizes and recognizes speech sounds using only acoustic cues, an articulatory system with a realistic vocal tract model, and a cognitive architecture which bridges the two. The environment in which the model resides is also simulated. The principal hypothesis guiding the model is that phonological development emerges from the interaction of auditory perception and hierarchical motor control. The model's auditory perception is specialized to segment and categorize speech into discrete phonetic events which closely correspond to discrete sets of functionally coordinated gestures learned by the vocal tract's articulatory apparatus. To imitate words, HABLAR need not solve the hard problem of relating continuous speech sound and continuous vocal tract motion. It learns the correspondence between one discrete sequence of events and another. To address both relative articulatory difficulty of different sounds and recombination of already mastered sounds into unique utterances, HABLAR views speech as a hierarchical control problem. An abstract phonological control level composes each utterance out of one or more elemental sounds, choosing which of many lower level articulatory controllers is most likely to generate each sound. With continuous proprioceptive feedback as a guide, the articulatory control level choreographs the fine-grained timing of vocal tract and pulmonary motions necessary to produce each elemental sound. Simulations of the model's auditory perception mimic categorical perception which develops in infancy. Motor control simulations demonstrate the feasibility of shaping multi-dimensional articulatory control with phonetic feedback. HABLAR and its properties qualitatively explain the phenomena of normal phonological development. It provides a computational testbed and a new framework with which to formulate and evaluate hypotheses about phonological development.