Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Glottal Waveform Synthesis with Volterra Shapers

Jean Schoentgen

Institute of Phonetics, Universite Libre de Bruxelles, Brussels, Belgium

Recently, we proposed a nonlinear zero-memory input-output model of the glottal pulse. This so-called Volterra model represents pulse shape with great accuracy. The model's driving function is a cosine. The number of model coefficients is typically equal to 2 x 40. Indeed, the model consists of two polynomials. The first models the even and the second the odd component of the glottis signal. The Volterra coefficients are directly calculated from the Fourier coefficients. The model outputs the original pulse when the amplitude of the driving cosine is equal to one, its phase is equal to zero and its frequency equal to the original frequency. In this article we suggest using the Volterra model for synthesizing glottal pulses. Models with a small number of control parameters are preferred for synthesis. Therefore, we use the Volterra model differently. We calculate once and for all the coefficients cj and dj for a small number of glottal pulses obtained from a speaker and we keep them fixed while we adjust amplitude, phase and frequency of the driving cosine. We thus generate glottal pulses of different shapes with the same model. Indeed, nonlinear systems are not characterized by a transfer function. This means that the shape of the output of a Volterra model depends on the shaping functions and also on control parameters of the driving cosine. We report an experiment that we carried out to test this synthesis method.

Full Paper

Bibliographic reference.  Schoentgen, Jean (1992): "Glottal waveform synthesis with volterra shapers", In ICSLP-1992, 1147-1150.