Parametric model of spectral envelope to synthesize realistic intensity variations in singing voice

In this paper, we propose a method to synthesize the natural variations of spectral envelope as intensity varies in singing voice. To this end, we propose a parametric model of spectral envelope based on novel 4-pole resonators as formant filters. This model has been used to analyse 60 vowels sung at different intensities in order to define a set of functions describing the global variations of parameters along intensity. These functions have been used to modify the intensity of 16 recorded vowels and 8 synthetic vowels generated with Vocaloid. The realism of the transformations performed with our approach has been evaluated by four amateur musicians in comparison to Melodyne for real sounds and to Vocaloid for synthetic sounds. The proposed approach has been proved to achieve more realistic sounds than Melodyne and Vocaloid, especially for loud-to-weak transformations.