Does swallowing bootstrap speech learning?

Abstract

Learning speech movements can be modeled as searching for sets of muscle activations in a high-dimensional space that satisfy task-specific criteria relevant to the language learner. An unstructured search of such a space is problematic: the high dimensionality makes non-heuristic search inefficient, and the number of redundant solutions for a given task makes predicting muscle activation difficult. Although most speech movements must be learned, humans can produce many of our most complex oral motor behaviors (swallowing, suckling, vocalizing, smiling, etc.) at birth. This indicates a degree to which the biomechanical and neural structures needed for complex action in the vocal tract appear to be built-in. A model is described, based on neurophysiological and computational motor research, in which action is governed by neuromuscular modules which can emerge and/or change ontogenetically through use. Simulations of tongue movements were conducted using a realistic 3D biomechanical model in ArtiSynth (www.artisynth.org; e.g., Stavness et al., 2012, Gick et al. 2014). Results of these simulations demonstrate that the neuromuscular activations leading to full oral swallowing closure are a subset of those that result in tongue bracing. This implies that submodules of the swallowing gesture can be recruited for use in tongue bracing, and makes a search of this space much simpler when the swallowing activation is used as a starting point. These findings are consistent with the view that phylogenetically encoded structures such as those needed for swallowing (e.g., MacNeilage 2008; Studdert-Kennedy & Goldstein 2003) bootstrap speech learning.