A Comparative Study of Various Quantization Schemes for Speech Encoding

Design of an efficient encoding scheme requires some knowledge of the statistics of the signal. Efforts to improve the performance of PCM systems have taken two primary directions : (i) Use of quantizing schemes based on knowledge of the (onedimensional) probability density function (PDF) of the samples to be quantized. (ii) Use of quantizing schemes exploiting the correlation between successive samples. If we had an a priori knowledge of the statistics of the samples, a nearly optimum quantization scheme could be used consisting of: (i) A quantizer matched to the PDF of the signal to be quantized. (ii) A predictor optimized for the given autocorrelation function of the signal. The predictor lowers the variance of the signal to be quantized by 1597 removing the correlation between successive samples. This is done by subtracting an estimation value from each incoming sample; the difference can be quantized, encoded, and transmitted (differential PCM = DPCM).