An investigation of vocal vibrato for synthesis

Abstract

High-fidelity singing synthesis requires careful consideration of the properties and character of natural vibrato. The research reported here evaluates some of the characteristics of vocal vibrato and presents a new panned-wavetable synthesis method. The singing tone 'a' was recorded digitally and analyzed for bass, tenor, alto and soprano voices, providing time-variant measurements of amplitude; frequency and phase for each of the voice partials. Formant 'tracings' provide a novel method for examining vocal spectra during vibrato. The significance of vibrato waveform parameters and the role of spectrum modulation (due to partial amplitude fluctuations) during the vibrato cycle was investigated by resynthesis of the singing tones from modified analysis data. Informal listeing to examples produced by the panned-wavetable synthesis model indicated that inclusion of typical random fluctuations of vibrato rate, vibrato depth and nominal sung frequency resulted in no quality preference over examples with constant values. Inclusion of vibrato-induced spectrum modulation resulted in a substantial improvement over examples having constant spectra.