UBC Theses and Dissertations
Adaptive predictive delta coder combining syllabic adaptation and a self-adaptive quantizer Schellenberg, Walter Ulrich
The purpose of this thesis is to evaluate an adaptive differential encoder for digital communication channels. The redundancy reduction technique used in this coder is not restricted to speech signals only. However, it was optimized for such signals with the objective of keeping the bit-rate as low as possible. In the transmitter, the signal redundancy is reduced in two steps. First, taking advantage of the quasi-periodicity of speech signals, the current signal value is predicted from the value one period before. Secondly, the difference between this prediction and the true value is estimated by a prediction based on the two previous differences. The error of this second prediction is quantized and transmitted to the receiver. The receiver produces a replica of the original signal by adding the received error signal to the predicted value. The adaptation of the quantizer step size has an exponential characteristic and contains a delay of one sampling period. These two features give rise to instabilities and poor reproduction of high signal frequencies, the latter being an inherent characteristic of this quantizer. The stability problem could not be solved theoretically because the nonlinearity and delay in the quantizer render the system mathematically intractable. By restricting the maximum quantizer level and adding a direct feedback of the step size to the second predictor, the instabilities are restricted to acceptable limits. The coder was simulated on a digital computer and optimized for sampling rates of 8, 12, and 16 kHz using objective calculations of the signal-to-quantization noise ratio as well as subjective preference tests. In some cases the calculated S/Q ratios allow no distinction in performance, but the subjective evaluations exhibit strong differences in preference. This observation emphasizes the necessity to examine the subjective performance of such voice systems. Comparisons with speech from a log PCM encoder indicate that at a sampling rate of 8 kHz, the subjective quality of the reconstructed speech is slightly superior to that of log PCM encoded at 3 bits per sample and the same sampling frequency. It is estimated that at a sampling rate of 8 kHz, an additional 3800 bits/sec are required to transmit the four coder parameters. The suggested final transmission bit-rate is therefore 11.8 kbits/sec. Thus, a data compression of approximately two to one has been achieved. Included are suggestions for reducing the bit-rate to 9.6 kbits/sec with minimal degradation in speech quality below that achieved using 11.8 kbits/sec.
Item Citations and Data