A study of voice source interaction
Mats Båvegård · 2009
INTRODUCTION Inverse filtering combined with extraction of LF-source parameters, e.g. Gobl(1988), Karlsson(1992) have provided a knowledge base for improved quality in speech synthesis, Carlsson, Granstrom and Karlsson (1990). A n inherent constraint is the linearity of LF-synthesis (Fant, Liljencrants and L in , 1985) which assumes independent source and vocal tract characteristics. In a more advanced model taking into account suband supraglottal interaction, e.g. that of L in (1990) the true glottal flow is dependent on the entire system. One of the interaction effects noted already by Ananthapadmanabha and Fant(1982), and by Fant and Lin(1987), Lin(1990) is the presence of ripple superimposed on the glottal flow. In the course of pursuing further work on complete articulatory synthesis we are interested in the perceptibility of glottal ripple and to what extent it adds to synthesis quality. According to earlier experiments, e.g. Nord et al. (1986) the perceptual effects are small. Our interactive synthesis system Tracttalk, (Lin, 1990) provides the facilities to perform more conclusive tests. One of the conditions we have tested is the combination of an L F shaping of glottal area with a constant leakage which provides interesting results. M E T H O D A N D RESULTS The LF-model (Fant et al., 1985) of glottal flow is here used as an area function of the glottal area. It is controlled with respect to three waveshape parameters and the fundamental frequency F 0 . These are the rise time parameter R g , the pulse asymmetry parameter R k and the duration of the residual closure phase T A . It is also possible to add a constant leak. This area function will then be used to compute the true glottal airflow of the interactive system (for computational details, see L i n , 1990). Fonetik -94 39 Two different voice types are modelled in this study. One symmetrical glottal area function, with different degrees of dynamic and constant glottal leakage and one with asymmetrical glottal area function and different amount of dynamic leakage. Figure 1. Two different area functions (af); first glottal flow pulse (1), second glottal flow pulse(2); first and second differentiated glottal flow pulses (1' & 2'). A: symmetrical area, B: asymmetrical area function, C: linear LF-glottal flow pulse with the same LF-parameters for the flow as for the area in B. Three different types of interaction effects can be shown from this experiment, see figure 1: Pulse skewing, due to the total vocal tract and glottal inductance. The glottal flow is skewed to the right compared to the glottal area. Superposition ripple, originating from short time variations of the transglottal pressure. This is evoked by formant oscillations of the pulse onset and previous glottal periods. Damping, of the flow oscillation in the open phase of the glottal pulse. These effects are common for this type of experiments and found in both the symmetrical and the asymmetrical area function simulations. More interesting is the simulations where a constant leak is introduced. Figure 2 shows simulations of the first formant of an /a/-vowel with a symmetrical L F area function where a constant leak of 0.03 cm 2 is established similar to a glottal chink. The area functions in figure 2, are modelled with T 0 = 8.0 ms, T p = 2.0 ms, T e = 4.0 ms and T A = 1.0 ms, which implies an R k value of 1.0. The constant lungpressure is 4.0 cm H 2 0 and the maximum glottal area is 0.15 cm 2 . 40 Working Papers 43, Dept of Linguistics and Phonetics, Lund, Sweden