The neural net model

the feature extraction and the learning algorithm for the speech recognition by neural nets are investigated in this paper. A non-uniformly windowed pyramidical architecture is proposed

the normalizations of the feature vectors of two extraction algorithms are investgated and the Polak-Ribiere learning algorithm is modified to guarantee the weight vector not to converge to non-stable local minima. Speaker-dependent experments show that over 99% accuracy can be achieved for recognizing five vowels and ten digits. Speaker-independent experiments involved 90 speakers from 25 provinces of China show that 90.3% accuracy can be obtained for the digit recognition

and the performance can be increased by 1.0% after the pre-processing window is employed. The windowed model is more applicable for speech recognitions.