Unsupervised learning of vocal tract sensory-motor synergies

The degrees of freedom problem is ubiquitous within motor control arising out of the redundancy inherent in motor systems and raises the question of how control actions are determined when there exist infinitely many ways to perform a task. Speech production is a complex motor control task and suffers from this problem, but it has not drawn the research attention that reaching movements or walking gaits have. Motivated by the use of dimensionality reduction algorithms in learning muscle synergies and perceptual primitives that reflect the structure in biological systems, an approach to learning sensory-motor synergies via dynamic factor analysis for control of a simulated vocal tract is presented here. This framework is shown to mirror the articulatory phonology model of speech production and evidence is provided that articulatory gestures arise from learning an optimal encoding of vocal tract dynamics. Broad phonetic categories are discovered within the low-dimensional factor space indicating that sensory-motor synergies will enable application of reinforcement learning to the problem of speech acquisition