Study of Control Strategy Mimicking Speech Motor Learning for a Physiological Articulatory Model

摘要

A full 3D physiological articulatory model was developed to simulate the mechanism of human speech production. In order to control the physiological articulatory model to generate speech sounds, we proposed a novel control strategy by constructing a mapping from acoustic objective to muscle activation patterns by simulating the babbling process in speech acquisition. This mapping was built up using a two-layer feedforward neural network. In order to evaluate the control strategy, muscle activation patterns were estimated from known speech sounds, and then were used to drive the physiological articulatory model to synthesize speech. The comparison results of the acoustic space and articulatory space showed that the proposed mapping-based control strategy was feasible for model control.