Singing Voice Synthesis Combining Excitation plus Resonance and Sinusoidal plus Residual Models

Abstract

This paper presents an approach to the modeling of the singing voice with a particular emphasis on the naturalness of the resulting synthetic voice. The underlying analysis/synthesis technique is based on the Spectral Modeling Synthesis (SMS) and a newly developed Excitation plus Resonance (EpR) model. With this approach a complete singing voice synthesizer is developed that generates a vocal melody out of the score and the phonetic transcription of a song.