Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence
This work concerns itself with the analysis of voiced speech signals, in particular the analysis of the glottal source signal. Following the source-filter theory of speech, the glottal signal is produced by the vibratory behaviour of the vocal folds and is modulated by the resonances of the vocal tract and radiation characteristic of the lips to form the speech signal. As it is thought that the glottal source signal contributes much of the non-linguistic and prosodical information to speech, it is useful to develop techniques which can estimate and parameterise this signal accurately. Because of vocal tract modulation, estimating the glottal source waveform from the speech signal is a blind deconvolution problem which necessarily makes assumptions about the characteristics of both the glottal source and vocal tract. A common assumption is that the glottal signal and/or vocal tract can be approximated by a parametric model. Other assumptions include the causality of the speech signal: the vocal tract is assumed to be a minimum phase system while the glottal source is assumed to exhibit mixed phase characteristics. However, as the literature review within this thesis will show, the error criteria utilised to determine the parameters are not robust to the conditions under which the speech signal is recorded, and are particularly degraded in the common scenario where low frequency phase distortion is introduced. Those that are robust to this type of distortion are not well suited to the analysis of real-world signals. This research proposes a voice-source estimation and parameterisation technique, called the Power-spectrum-based determination of the Rd parameter (PowRd) method. Illustrated by theory and demonstrated by experiment, the new technique is robust to the time placement of the analysis frame and phase issues that are generally encountered during recording. The method assumes that the derivative glottal flow signal is approximated by the transformed Liljencrants-Fant model and that the vocal tract can be represented by an all-pole filter. Unlike many existing glottal source estimation methods, the PowRd method employs a new error criterion to optimise the parameters which is also suitable to determine the optimal vocal-tract filter order. In addition to the issue of glottal source parameterisation, nonlinear phase recording conditions can also adversely affect the results of other speech processing tasks such as the estimation of the instant of glottal closure. In this thesis, a new glottal closing instant estimation algorithm is proposed which incorporates elements from the state-of-the-art techniques and is specifically designed for operation upon speech recorded under nonlinear phase conditions. The new method, called the Fundamental RESidual Search or FRESS algorithm, is shown to estimate the glottal closing instant of voiced speech with superior precision and comparable accuracy as other existing methods over a large database of real speech signals under real and simulated recording conditions. An application of the proposed glottal source parameterisation method and glottal closing instant detection algorithm is a system which can analyse and re-synthesise voiced speech signals. This thesis describes perceptual experiments which show that, iunder linear and nonlinear recording conditions, the system produces synthetic speech which is generally preferred to speech synthesised based upon a state-of-the-art timedomain- based parameterisation technique. In sum, this work represents a movement towards flexible and robust voice-source analysis, with potential for a wide range of applications including speech analysis, modification and synthesis.
O Cinneide, A. (2012) Phase-Distortion-Robust-Voice-Source Analysis. Doctoral Thesis. Dublin, Technological University Dublin. doi:10.21427/D7FW3B