Designing and Controlling a Source-filter Model for Naturalistic and Expressive singing voice synthesis

In this paper, we describe a voice synthesis model developed for musical purposes. Based on a source-filter model, this synthesizer has been specifically designed to allow the synthesis of natural sounding singing voices by including pitch and amplitude variations and by careful tuning of consonant to vowel transitions. A particular attention is given to the reproduction of plosive consonants. The model covers all singing voice registers, from bass to soprano, and allows the control of several tone quality parameters such as vibrato depth and frequency, voice roughness and articulation speed. Its database is structured to synthesize whole consonant-vowel syllables. As a result, it is relatively easy to construct musically expressive phrases with just a few manipulations and control commands. The model uses Csound as the audio engine and can produce several voices at small cost in CPU.