Simple neural networks for speech research
M.P. Beddoes · 2002
The paper is about a model of a physical system, the lips, throat and glottis, used in speech production. Simple steps are outlined for getting a neural network to model the system. The approach is illustrated by examples. An interesting example is a two tube model of the throat. Starting with an "electrical engineering model" of the glottis in the form of the Kelly-Lochbaum equations, an Overall Signal Flow diagram, we get: (a) the neural network layout directly; (b) the weights between various neural processing elements. The problem of designing the layout topology for optimum performance neural network has received much attention. Hence the importance of (a) which guarantees a network with the precisely the same characteristics as the electrical engineering model. If the electrical engineering model is optimum, then the derived ANN will be optimum. The model will give the steady state output wave resulting from a steady state glottal wave. The glottal wave can be obtained by a glottograph or by photoglottography. One can choose the glottograph because it interferes little with the speech process. The glottal wave thus obtained can test the model's ability to produce the corresponding speech wave. The N section tube model is lossless. From this it follows that it can be inverted to give the glottal wave given the speech wave. The inverted network may be able to act as a tool to investigate whether there is a reliable correlation between emotion of the speaker, such as fear, anger, even love, and the glottal wave.