Effective modeling of sounds and voice is of great importance to provide a
realistic and expressive acoustic layer in interactive environments,
multimodal interfaces and communication systems.
The talk presents some recent advances concerning the use of physical
modeling and machine learning methods in sound and voice synthesis. We
also discuss an approach to the modeling of acoustic systems that combines
prior information, exploited through physical modeling, and nonlinear
dynamics reconstruction, exploited through kernel regression modeling. We
demonstrate this approach on some particular case studies, including
everyday sounds and voice phonation.