Speech coding using spline wavelet and adaptive backward prediction

Abstract

Wavelet techniques are well known for providing both time and frequency localization for signal processing. Although great achievements in wavelet applications have been obtained in fields such as signal detection and image compression, research effort seems to be lacking in the application of using semi-orthogonal wavelets to speech coding, especially in real-time mode. In this thesis, the application of splinewavelets incorporating adaptive backward prediction to speech coding is investigated. The goal of the research is to develop the real-time, moderately complex algorithms which take advantage of the multiresolution analysis generated by the wavelet transform in order to provide a high-quality speech signal at a total bit rate of 16 Kbps and intelligible reproduction at a bit rate of 8 Kbps. To achieve this, the existing wavelet algorithms are modified to fit the requirements of real-time data processing. Furthermore, since the input to the coder are wavelet coefficients which are. quite different from speech signals, a new nonuniform adaptive quantizer algorithm based on the pdf-optimized quantizer design is introduced. Several comparisons are given to show that the new design is better than the tradictional Jayant's adaptive quantizer (JAQ). Combining it with the autocorrelation solution predictor, which is redesigned to work on the backward mode in this research, the applied ADPCM (Adaptive Differential Pulse Code Modulation) algorithm is proven to effectively encode the wavelet coefficients. Finally, the coding system containing both wavelet analysis and the developed ADPCM is built up, and various simulations are presented to show the advantages of applying wavelet techniques in speech coding.