Abstract:
Under normal conditions, the vocal tract shapes (postures) used by speakers to produce vowels appear to represent only a small subset of the possible shapes that could be used. This seems to suggest that speakers use postural targets (e.g., constriction targets) when planning speech movements. However, postural targets limit the motor equivalent capabilities of a movement controller. An alternative explanation for the approximate postural invariance maintains a high degree of motor equivalence by utilizing only acoustic/perceptual targets and transforming trajectories planned in acoustic/perceptual space into articulator movements through a directional mapping with a ``postural relaxation'' component. The postural relaxation component biases movements toward more central articulator configurations. Examples of this kind of controller include robotics techniques that utilize a Jacobian pseudoinverse in combination with null space movements formed using an orthogonal projection operator. The current work investigates how such a directional mapping can be learned by a biological motor system during babbling and how it can lead to postural invariance. Simulations of the DIVA model of speech acquisition and production utilizing an adaptive postural relaxation scheme verify the property of approximate postural invariance despite using no postural targets in the planning process. [Work supported by NIH and the Sloan Foundation.]