Generating a phonetic transcription of the speech waveform is one method which can be applied to continuous speech recognition. Current methods of labelling a speech wave involve the use of techniques based on spectrographic analysis. This paper presents a computationally simple method by which some phonemes can be identified primarily by their shapes. Three shapes which are regularly manifested by three phonemes were examined in utterances made by a number of speakers. Features were then devised to recognise their patterns using finite state automata combined with a checking mechanism. These were implemented in the Wave Analysis Language (WAL) system developed at the University of Canberra and the results showed that the phonemes can be recognised with high accuracy. The resulting shape features have also demonstrated a degree of speaker independence and context dependency.

Concepts of tree coding and of rate-distortion theory are applied to the problem of the transmission of analog signals over digital channels. Coding schemes are developed which yield improvements of up to six dB in signal-to-noise ratio over conventional techniques for the reproduction of speech waveforms. / Thesis / Master of Engineering (MEngr)