Speech coding at low to medium bit rates.

Abstract

The research reported herein is concerned with the development of high quality speech coding algorithms at bit rates between 4 and 8 kb/s. In the literature, analysis by synthesis coding based on a weighted mean squared error distortion criteria has been demonstrated to work well at rates as low as 4-16 kb/s. In this research, improved search techniques coupled with improved codebook design methodologies are proposed to improve the performance of conventional code-excited linear predictive coders. Improved methods for quantizing the short term filter are developed by applying a tree search algorithm and joint codebook design to multi-stage vector quantization. Joint codebook design procedures are developed to design locally optimal multi-stage codebooks. Weighting during centroid computation is introduced in order to improve the outlier performance of the multi-stage vector quantizer. Multi-stage vector quantization is shown to be both robust against input characteristics and in the presence of channel errors. Spectral distortions of approximately 1 dB are obtained at rates between 22 and 28 bits/frame for various multi-stage configurations. Structured codebook design procedures for the excitation in code-excited linear predictive coders are compared to general codebook design procedures. It is determined that little is lost in using significant structure in the excitation codebooks, while greatly reducing the search complexity. Sparse multi-stage configurations are proposed as a means of reducing computational complexity and memory size while offering both a variable rate and variable complexity. Improved search procedures applied to code-excited linear prediction which attempt to jointly optimize the short term filter, the adaptive codebook, and the excitation are proposed. Although the computational complexity is rather large, segmental signal to noise ratio improvements between 1 and 2 dB are realized in practice.