Estimation of articulatory motion using neural networks

This paper discusses an application of neural networks to the problem of estimating articulatory parameters from speech waves. Since three- or four-layer feed-forward neural networks can approximate arbitrary nonlinear functions, and since estimation of articulatory parameters can be regarded as nonlinear mapping, it is expected that the neural network can be applied to the articulatory parameter estimation problem. Our results show that using a neural network, the articulatory parameters are estimated faster and more accurately than with the conventional model matching method.