Date of Award
Doctor of Engineering Science in Electrical Engineering
W. H. Warren Ball
Frederick A. Russell
The objective of the research reported here is the design of efficient speech coders that can easily be implemented in integrated circuit hardware. Companding techniques like those introduced by M. R. Winkler, J. A. Greefkes, F. DeJager, A. Tomozawa and H. Kaneko were explored along with a large body of theory concerning the application of linear prediction to speech coding.
The best features of the speech signal to be measured and coded are the overall amplitude, the resonant frequencies and dampings of the vocal cavity and the fundamental frequency of the vocal cord oscillations. Adaptive quantization was used to track variations in overall amplitude, and adaptive prediction was used to track the frequencies and dampings of the cavity resonances. No attempt was made to exploit redundancies related to the vocal cord oscillations, however.
An adaptive differential pulse code modulator (i.e., an ADPCM coder) with a fixed integrator was simulated first. Later a hardware model was constructed, signal to noise measurements were taken and subjective tests conducted. When operating at 4 bits per sample, speech of a quality nearly equal to that of 7 bit log PCM was regenerated by the ADPCM encoder. At 3 bits per sample speech quality was nearly equal to 6 bit log PCM.
Further improvements were achieved with the application of adaptive predictors in place of the integrator. The predictor coefficients form a vector which is adapted in a direction away from the gradient with respect to the error power. By applying this technique to the quantized signals occurring in the coder, the coefficients are derived from the quantized error signal; hence, there is no need to transmit them.
Cummiskey, Peter, "Adaptive, differential pulse-code modulation for speech processing" (1973). Dissertations. 1287.