The Digital Signal Processing Lab @ UCSD

The Digital Signal Processing Lab @ UCSD

Thesis Title

Modeling and Quantization Techniques for Speech Compression Systems

Thesis Abstract

This dissertation introduces new modeling and quantization techniques for improving the performance of speech compression systems.

In Part I, mixed-phase all-pole models are proposed for modeling the spectral magnitude and phase information in voiced speech. Mathematical and physiological motivations for the models are given.

Time domain and frequency domain parameter estimation algorithms are described, including efficient least-squares based algorithms and algorithms which incorporate perceptual masking effects. It is shown that high quality voiced speech can be reconstructed from a fourteenth order mixed-phase all-pole model.

In Part II, vector quantization of the filter coefficients of all-pole models is investigated. A theoretical analysis of high rate vector quantizers trained using suboptimal distortion measures is presented. It is shown that, at high rate, the quantization distortion approaches a quadratically weighted error measure. The quadratic weighting matrix is a "sensitivity matrix," which is a generalization of the scalar sensitivity concept. The diagonal elements of the sensitivity matrix relate to the sensitivities of the individual scalar parameters, and the off-diagonal elements relate to the interactions which occur when multiple parameters are quantized simultaneously. The sensitivity matrices, with respect to the Log Spectral Distortion measure, for standard linear predictive coding (LPC) filter coefficients, reflection coefficients, log area ratios, arcsine coefficients, and line spectral pair (LSP) frequencies are described. In particular, it is shown that the sensitivity matrix for the LSP frequencies is diagonal, implying that a vector quantizer trained by minimizing a weighted mean squared error measure in the LSP frequencies can provide optimal performance at high rate. Computationally efficient algorithms for computing the sensitivity matrices are described. Finally, the theoretical results are compared with experimental data and are shown to be valid for quantization rates of interest.

Year of Graduation: 1994