Abstract
Since the very early years of this century, communication systems have been subject to unparalleled growth in both capacity and network diversity. Voice communications, an ever dominant section of the market, has largely motivated the development of many new applications such as ISDN and Digital Mobile Radio. With communications infrastructure increasingly switching over to digital systems, digital voice coding has been subject to feverish development in recent years. Vocoders, historically among the first speech compression systems, received great initial interest due to their simple implementation. Although operating below 4.8kbit/s, these simple early vocoders were not capable of the quality demanded of public applications and therefore remained a niche product; mostly applied in private networks. Due to the importance of commercial public communications, vocoders have been relegated to the background of mainstream research in the last decade by highly complex analysis-by-synthesis algorithms operating at rates between 8kbit/s and 16kbit/s. However, with the quality of these systems apparently saturating at 6kbit/s, the application of very powerful digital processors to improved vocoder implementations has recently seen greater interest. This thesis reports on investigations into very low bit rate vocoder techniques, operating in the range of 2.4kbit/s to 4.8kbit/s. Previous research into vocoders has clearly identified the accurate representation of the speech excitation signal as the greatest obstacle to high quality. The hybrid coders that have dominated the last decade are largely a result of this fact. While the characteristics of spectral envelope modelling in vocoders is discussed in this thesis, most effort has been directed at the excitation problem. The characteristics of human speech excitation are introduced along with the critical problems of pitch determination and synthetic excitation generation. Two main vocoder studies are reported, both concentrating upon proposed excitation improvements to source-filter based vocoders. The first coder investigates a purely analytical encoding of the speech excitation signal using cepstral analysis. The second project attempts to improve a channel vocoder by spectrally mixing voiced and unvoiced synthetic excitations, to better resemble the true mixed excitation nature of speech.