Comparing the hybrid HMM system with discrete ones, the relative error rate can be reduced by about 40% using the neural network vector quantizer instead of the k-means VQ for on-line recognition and by about 20% for off-line handwriting recognition.