The autocorrelation function provides basic data for speech signal processing by the linear prediction algorithm. The first step toward developing a practical speech signal processing technique is noise reduction in the autocorrelation function of speech corrupted by noise. Recent developments in neural network (NN) techniques have achieved a respectable performance of classifiers even in the worse condition where SNR is less than 0 dB. The vector quantization (VQ) technique, on the other hand, also provides reasonable bases for the discrete description of continuous speech signals. The combination of the two techniques gives the possibility of excellent noise reduction in speech signal processing. Some recent experimental results based on the above principle are given in detail.