Pole-zero modeling of speech

Abstract

In this dissertation, pole-zero modeling of clean and noisy speech has been investigated. The purpose of this work is to improve the quality of the conventional linear prediction vocoder by studying the following two aspects : accurate representation of spectral valleys, especially those of nasal or consonant sound, and accurate spectral estimation of speech corrupted by noise. Four types of polezero modeling methods have been studied. They are the method of modified Yule-Walker (MYW) with time domain inverse linear prediction, the modified least square (MLS) method, the method of modified least square with autocorrelation compensation (MLSAC) and the autocorrelation prediction (AP) method. Of those, the first three have been newly proposed. These four algorithms have been derived in a unified approach that is based on high-order pole model fitting and decomposition method. Since these algorithms need only linear operations to get solutions, they are computationally more efficient than any other pole-zero modeling methods that usually resort to iterative algorithms. Those algorithms studied require far less computations than other previously proposed iterative algorithms. When these algorithms are used, one needs to have only 4 to 7 times more computations than that required for the all-pole analysis. The MYW method derived based on the AP method yields accurate spectral estimate for nasal sound. Also, it yields the most accurate spectral estimate for noisy synthetic speech than the other methods. However, when noisy natural speech is used as the input signal, the method does not yield a spectral estimate any better than the all-pole method. The MLS method that can be regarded as a subclass of the MLSAC method is similar in concept to those methods studied by Kalman, and Mullis and Roberts. Although the MLS method is computationally efficient, it appears that the performance improvement over the all-pole modeling method is not significant. On the other hand, ML...