Glottal waveform synthesis with Volterra shaping functions

We recently proposed an input-output model of the glottal pulse. Mathematically speaking, the pulse is broken down into a cosinusoidal input signal and a pair of nonlinear shaping functions. The pulse is recovered when the cosinusoid is put through the shapers. In this article, it is shown that the cycles of a speaker's glottal waveform can be synthesized with the shaping functions of a small number of reference cycles. Indeed, nonlinear systems are not described by a transfer function. Therefore, it may be assumed that the nonlinear shaping functions of a glottal pulse are less variable than the shape of the pulse itself. Two experiments were carried out to test this assumption. In a first, the output static waveforms from a two-mass model of the vocal folds were copied. In a second, the glottis signal that was obtained from a logatome [ama] spoken by a male speaker was analyzed and synthesized. Each pulse was characterized by its peak amplitude, period and form factor. In both experiments, the features of all the glottal pulses could be copied by calculating the shaper coefficients of just two reference pulse and by adjusting the control parameters of the driving cosinusoid till the output of the shaper exhibited the desired feature values. © 1992.SCOPUS: ar.j