Contributions to recognition and low bit rate coding of speech
Thesis DisciplineElectrical Engineering
Degree GrantorUniversity of Canterbury
Degree NameMaster of Engineering
Original research contributions to aspects of automatic recognition and low bit rate coding of speech are presented. A comprehensive review of the historical development of speech processing is provided. Various facets of the physiology of the human speech production organs, the classifications of speech sounds, and human perception of speech sounds are discussed. Essential mathematical. theories for speech processing are outlined. Three different techniques for automatic speech recognition are discussed in detail. One of these techniques is implemented, modified and extended. It is found that these modifications and extensions result in improved recognition rate of the digits zero to nine uttered by several New Zealand speakers. These results are comparable to those of studies conducted elsewhere. A novel speech coding scheme is examined. It is found that the scheme produces good quality speech at 16 kbps. However, there are some unpleasant click sounds which are believed to be caused by the instability of the scheme. The scheme is modified by adding an impulse function to the glottal pulse. This is found to be effective in stabilising the scheme and in removing the click sounds. Because errors are inevitable in any speech coding scheme, it is important to quantify these errors and to analyse their nature, as the analysis may provide useful hints on ways in which these errors can be eliminated. To this end, the reconstructed speech is assessed alongside two reference signals, which are formed by adding known amounts of multiplicative and additive noise, respectively, to the original speech. These assessments strongly suggest that the errors in the reconstructed speech are largely caused by the multiplicative noise.