Complicated and risky works can be done effectively by a robot controlling using human voice. For successful application of voice recognition technique to robot control, a simple recognition method with short computational time is required and the number of words has to be selected suitably. In this thesis, the problem of voice recognition for robot control is considered. The voice recognition method uses multilayed neural networks using log vocal tract area ratio as for the selected words recognition. The estimation method uses a adaptive inverse filters to autocorrelation coefficients. The network learning algorithm described here is based on the conjugate gradient method. In order for a robot not to move when any unknown word to the network is spoken, an algorithm is proposed. The proposed algorithm used the total output neuron error of neural networks. If the total error of a word is more than threshold, the word is rejected.
