Phase Estimation with Application to Speech Analysis-Synthesis

摘要

This dissertation addresses the problem of estimating the phase of the frequency response of mixed phase signals and systems. A number of techniques are applied to estimation of the phase frequency response of the speech production tract from quasi-periodic speech segments. Methods of phase estimation are categorized as indirect or direct. A subject of the indirect procedures yield a closed form solution for retrieving the phase from the magnitude of a mixed phase frequency response and a priori knowledge about the corresponding signal. Linear iterative algorithms are also developed for retrieving the phase from the magnitude, and, similarly, the magnitude from the phase, with a casuality or finite duration constraint imposed on the desired signal. The iterative algorithm for magnitude retrieval provides an alternative to the Hilbert transform for obtaining the magnitude from the phase of a minimum phase signal, but without the need of an unwrapped phase. In addition, it serves as the major component within a new phase unwrapping algorithm which does not require modulo 2(pi) considerations. An alternate indirect strategy changes a phase estimation problem to one of magnitude estimation by modifying a quasi-periodic waveform so that the desired impulse response takes on a minimum phase characteristic.