In this paper, the performance of neural network for isolated word recognition based on the movement of the mouth is investigated. Ten isolated words "zero" to "nine" form the vocabulary. The three parameters: height between the upper and lower lips, width of the mouth from left edge to right edge and the specific angle of the upper lip, are selected as representative features of the shape of the mouth. A oneagainst-the-rest neural network is used for classification. The architecture of the neural network consists of ten sub-units. Each sub-unit distinguishes 1 word from the other 9 and has 50 input vectors, 10 hidden nodes and 1 output neuron. By suitably combining the outcomes from the 3 features individually, 92.5% accuracy of recognition may be attained.

Accepted version