A new voice source model based on high-speed imaging and its application to voice source estimation

Abstract

There are numerous models of varying complexities which seek to efficiently represent the voice source signal. These models are typically based on data and observations which can come from air-flow masks, electroglottographs, mechanical systems, and the inversefiltering of speech signals. The first part of this study examines observations from the high-speed imaging of the larynx and proposes a new source model, which is shown to provide a better fit for the observed data than existing models. The proposed source model is then used in an automatic source estimation application, based on methods introduced in an earlier study [1]. Results, on average, show that the proposed model provides a more accurate estimation of the source signal compared with the Liljencrants-Fant model.

Variability in the relationships among voice quality, harmonic amplitudes, open quotient, and glottal area waveform shape in sustained phonation

“…Studies to date suggest that the relationship between H1-H2 and OQ may be variable, but model fit to empirical pulse shapes is not such that definitive conclusions can be drawn. Because existing empirical data are not sufficient to clarify this situation, this study examined the relationship between H1*-H2* (measured from recorded acoustic signals), OQ, and the asymmetry coefficient (the length of the opening phase relative to the open phase, e.g., Henrich et al, 2001;Shue and Alwan, 2010), 2 measured synchronously from high-speed video images of the vibrating vocal folds. Note that previous work on this topic has used models of the glottal flow, which may differ in pulse skewness from the glottal area functions measured here (e.g., Howe and McGowan, 2007).…”

Perceptual evaluation of voice source models

“…The LF model (Fant et al, 1985) combines an exponentially increasing sinusoidal function and an exponential function with one amplitude parameter (E e ) and three time points; the Fujisaki-Ljungqvist model (Fujisaki and Ljungqvist, 1986) uses polynomials to model the shape and duration of different segments of the flow derivative waveform. Recent studies (Shue and Alwan, 2010;Chen et al, 2012) have proposed models of the glottal area waveform (as derived from high-speed endoscopic recordings of the laryngeal vibrations) rather than the flow pulse or its derivative. With four parameters, the first of these (Shue and Alwan, 2010) uses a combination of sinusoidal and exponential functions similar to the LF model, but with the ability to adjust the slopes of the opening and closing phases separately.…”

“…This paper assesses the relationship between model fit and perceptual accuracy by studying five time-domain source models (three of which are related)-the Rosenberg model (Rosenberg, 1971), the Fujisaki-Ljungqvist model (Fujisaki and Ljungqvist, 1986), the Liljencrants-Fant (LF) model (Fant et al, 1985), and two models proposed by Alwan and colleagues (Shue and Alwan, 2010;Chen et al, 2012)-and one model that describes the voice source in the spectral domain (Table I; Kreiman et al, 2014;see also Cummings and Clements, 1995). The Rosenberg model (Rosenberg, 1971), in contrast to the other models, describes the opening and closing phases of the glottal flow volume velocity with separate trigonometric functions that incorporate two timing parameters and one amplitude parameter.…”

“…Mathematical models of the voice source have been designed to provide high quality voicing for synthetic speech while minimizing the bandwidth for its transmission, and to model perceptually important aspects of the voice source. In this study we assessed the fit of five such models of the voice source-the Rosenberg model (Rosenberg, 1971), LiljencrantsFant (LF) model (Fant et al, 1985), Fujisaki-Ljungqvist model (Fujisaki and Ljungqvist, 1986), and two models proposed by Alwan and colleagues (Shue and Alwan, 2010;Chen et al, 2012)-to the shape of the glottal flow derivative and to glottal source spectra, and then examined the ability of the models to match the target voice qualities.…”

“…As presently implemented, most models of the voice source describe time-domain features of vocal fold vibration or of glottal flow, including peak glottal opening/peak flow and the negative peak of the flow derivative, which is associated with the instant of maximum excitation (Fant, 1993). Models are typically evaluated in terms of their fit to empirical pulse shapes (e.g., Fujisaki and Ljungqvist, 1986;Gobl and N ı Chasaide, 2010;Shue and Alwan, 2010). For example, the LF model captures changes in the glottal flow derivative using sinusoidal and exponential functions (Fant et al, 1985), and Fant attributed its widespread use to its ability to capture "essentials" of a variety of glottal wave shapes (Fant, 1995, p. 119).…”