PURPOSE: To improve the precision of word recognition by obtaining input values to a neural network which correspond to the time expansion and contraction of a voice caused by the difference of phonation habits among habits among people and phonation dispersion of the person who is making the phonation.

CONSTITUTION: An input voice is divided into prescribed time frames, obtain the power of the input voice for every frame, detect the frame which corresponds to the maximum point of the power time series which are obtained, compute the average value of the frequency characteristics within the frame which corresponds to the maximum point, compute the average values of the frequency characteristics of the frames which are located before and after the frame corresponding to the maximum point, compare the difference value and the threshold value which is set beforehand, divide the input voice using the frame, at which the difference value exceeds the threshold, as the boundary and use these average values of the frequency characteristics of the input voice for every block, which is divided, as inputs to the neural network 6.