On the use of voice descriptors for glottal source shape parameter estimation

This paper summarizes the results of our investigations into estimating the shape of the glottal excitation source from speech signals. We employ the Liljencrants-Fant (LF) model describing the glottal flow and its derivative. The one-dimensional glottal source shape parameter R-d describes the transition in voice quality from a tense to a breathy voice. The parameter R-d has been derived from a statistical regression of the R waveshape parameters which parameterize the LF model. First, we introduce a variant of our recently proposed adaptation and range extension of the R-d parameter regression. Secondly, we discuss in detail the aspects of estimating the glottal source shape parameter R-d using the phase minimization paradigm. Based on the analysis of a large number of speech signals we describe the major conditions that are likely to result in erroneous R-d estimates. Based on these findings we investigate into means to increase the robustness of the R-d parameter estimation. We use Viterbi smoothing to suppress unnatural jumps of the estimated R-d parameter contours within short time segments. Additionally, we propose to steer the Viterbi algorithm by exploiting the covariation of other voice descriptors to improve Viterbi smoothing. The novel Viterbi steering is based on a Gaussian Mixture Model (GMM) that represents the joint density of the voice descriptors and the Open Quotient (OQ) estimated from corresponding electroglottographic (EGG) signals. A conversion function derived from the mixture model predicts OQ from the voice descriptors. Converted to R-d it defines an additional prior probability to adapt the partial probabilities of the Viterbi algorithm accordingly. Finally, we evaluate the performances of the phase minimization based methods using both variants to adapt and extent the R-d regression on one synthetic test set as well as in combination with Viterbi smoothing and each variant of the novel Viterbi steering on one test set of natural speech. The experimental findings exhibit improvements for both Viterbi approaches. (C) 2013 Elsevier Ltd. All rights reserved.