A Self-Organizing Neural Model for Motor Equivalent Phoneme Production

This paper describes a model of speech production called DIVA that highlights issues of self-organization and motor equivalent production of phonological units. The model uses a circular reaction strategy to learn two mappings between three levels of representation. Data on the plasticity of phonemic perceptual boundaries motivates a learned mapping between phoneme representations and vocal tract variables. A second mapping between vocal tract variables and articulator movements is also learned. To achieve the flexible control made possible by the redundancy of this mapping, desired directions in vocal tract configuration space are mapped into articulator velocity commands. Because each vocal tract direction cell learns to activate several articulator velocities during babbling, the model provides a natural account of the formation of coordinative structures. Model simulations show automatic compensation for unexpected constraints despite no previous experience or learning under these constraints. Overview of the DIVA model Production of an acoustic signal that invariantly conveys to listeners a particular phoneme is carried out with large variability in articulator movements from one instance to the next (e.g., [1],[3],[5]). The process of producing an invariant result in a motor system despite large variations in the contributions of individual components from trial to trial is called motor equivalence. This paper investigates the relationship between phonological units and the motor actions that realize them, focussing on how motor plans to produce phonemes can be learned, and how these plans can lead to motor equivalent phoneme production. Figure 1a is a block diagram of the current model, which is named DIVA because a key component is a transformation from Directions (in vocal tract configuration space) Into Velocities of Articulators. This model assumes the existence of a neural representation of phonological units, or phoneme map, for production of speech. These units are each associated with a plan consisting of a set of synaptic weights that encode target values of vocal tract variables determining key acoustic properties of the vocal tract. These plans are then converted into appropriate articulator movements. The chosen articulator movements depend on context and external conditions, thus resulting in motor variability as seen in human speech. In addition, DIVA assumes the existence of a phoneme recognition system that transforms an appropriate incoming acoustic signal into a representation of the corresponding phonological unit. The model contains two learned mappings, indicated by filled semicircles in the figure: a mapping between phoneme representations and corresponding vocal tract configurations (or motor plans), and a mapping between vocal tract variables and articulator movements that realize desired vocal tract configurations. The levels of representation in DIVA are very similar to those of the speech production model of [6]. However, the nature of the mappings between between these levels differ. Furthermore, mappings in (6] were predefined by the modelers. In DIVA, emphasis is placed on how these mappings can self-organize; that is, how can an infant's speech production system learn the parameters governing this complex dynamical system through self-generated babbling witho1tt an external teacher? Mapping from phoneme representations to vocal tract configurations The variability in articulator positions seen during production of the same phoneme speaks against the explicit control of the spatial positions of speech articulators. Instead, direct, invariant control of higherlevel variables such as bilabial separation or tongue body constriction seems to be used. This result is not surprising; whereas these higher-level variables, or vocal tract variables, directly correspond to the acoustic properties of the vocal tract, the effect of an individual articulator is dependent upon the locations of other articulators. An efficient controller should utilize the flexibility afforded by the redundant set of articulators to invariantly produce acoustic information under a variety of circumstances. 'This work was supported in part by grants NSF IRI-87-16960 and NSF IRI-90-24877. ~~"~om~ lei Phoneme Vocal Tract ~--·· Map Conflg· uratlon