Voice Source Model Parameters from Spectral Domain

Voice source (or glottal pulse) parameters are usually specified in the time domain and are derived from inverse filtering. However, time domain waveform is extremely sensitive to recording conditions, highpass filtering and phase response of recording channel. Also, source-filter interaction will affect the waveform. In this paper a frequency (spectral) domain method of estimating the parameters of a voice source model is proposed. A systematic study of model generated pulses is conducted. It is shown that the frequency of the spectral peak is related to the open quotient. The assymmetry of the spectral peak in terms of 3 dB points is related to speed quotient. The nonabrupt glottal closure is known to be realted to special tilt. Voice source parameters are obtained from the spectrum both by direct measurement and also by an optimization scheme.