Differential encoding techniques applied to speech signals

The increasing use of digital communication systems has produced a continuous search for efficient methods of speech encoding. This thesis describes investigations of novel differential encoding systems. Initially Linear First Order DPCM systems employing a simple delayed encoding algorithm are examined. The systems detect an overload condition in the encoder, and through a simple algorithm reduce the overload noise at the expense of some increase in the quantization (granular) noise. The signal-to-noise ratio (snr) performance of such d codec has 1 to 2 dB's advantage compared to the First Order Linear DPCM system. In order to obtain a large improvement in snr the high correlation between successive pitch periods as well as the correlation between successive samples in the voiced speech waveform is exploited. A system called "Pitch Synchronous First Order DPCM" (PSFOD) has been developed. Here the difference Sequence formed between the samples of the input sequence in the current pitch period and the samples of the stored decoded sequence from the previous pitch period are encoded. This difference sequence has a smaller dynamic range than the original input speech sequence enabling a quantizer with better resolution to be used for the same transmission bit rate. The snr is increased by 6 dB compared with the peak snr of a First Order DPCM codea. A development of the PSFOD system called a Pitch Synchronous Differential Predictive Encoding system (PSDPE) is next investigated. The principle of its operation is to predict the next sample in the voiced-speech waveform, and form the prediction error which is then subtracted from the corresponding decoded prediction error in the previous pitch period. The difference is then encoded and transmitted. The improvement in snr is approximately 8 dB compared to an ADPCM codea, when the PSDPE system uses an adaptive PCM encoder. The snr of the system increases further when the efficiency of the predictors used improve. However, the performance of a predictor in any differential system is closely related to the quantizer used. The better the quantization the more information is available to the predictor and the better the prediction of the incoming speech samples. This leads automatically to the investigation in techniques of efficient quantization. A novel adaptive quantization technique called Dynamic Ratio quantizer (DRQ) is then considered and its theory presented. The quantizer uses an adaptive non-linear element which transforms the input samples of any amplitude to samples within a defined amplitude range. A fixed uniform quantizer quantizes the transformed signal. The snr for this quantizer is almost constant over a range of input power limited in practice by the dynamia range of the adaptive non-linear element, and it is 2 to 3 dB's better than the snr of a One Word Memory adaptive quantizer. Digital computer simulation techniques have been used widely in the above investigations and provide the necessary experimental flexibility. Their use is described in the text