Prediction and discrimination in neural networks for continuous speech recognition