Sound synthesis for physics-based computer animation.

摘要

In this thesis, we explore the problem of synthesizing realistic soundtracks for physics-based computer animations. While the problem of producing realistic animations of physical phenomena has received much attention over the last few decades, comparatively little attention has been devoted to the problem of generating synchronized soundtracks for these simulations. Recent work on sound synthesis in the computer graphics community has largely focused on producing sound for simple, rigid-body animations. While these methods have been successful for certain scenes, the range of examples for which they produce convincing results is quite limited. In this thesis, we introduce a variety of new sound synthesis algorithms suitable for generating physics-based animation soundtracks. We demonstrate synthesis results on a variety of animated scenes for which prior methods are incapable of producing plausible sounds.;First, we introduce a new algorithm for synthesizing sound due to nonlinear vibrations in thin shell structures. Our contributions include a new thin shell-based dimensional model reduction approach for efficiently simulating thin shell vibrations. We also provide a novel data-driven model for acoustic transfer due to vibrating objects, allowing for very fast sound synthesis once object vibrations are known. We find that this sound synthesis method produces significantly more realistic results than prior rigid-body sound synthesis algorithms for a variety of familiar objects.;Next, we further address the limitations of prior sound synthesis techniques by introducing a new method for synthesizing rigid-body acceleration noise—sound produced when an object experiences rapid rigid-body acceleration. We develop an efficient impulse-based model for synthesizing sound due to arbitrary rigid-body accelerations and build a system for modeling plausible rigid-body accelerations due to contact events in a standard rigid-body dynamics solver. This allows us to efficiently recover acceleration sound using data readily available from rigid-body simulations. Our results demonstrate that our method significantly improves upon the results available when using traditional rigid-body sound synthesis with no acceleration noise modeling. We also introduce a scalable proxy model which provides us with a practical method for synthesizing acceleration sound from scenes with hundreds to thousands of unique objects. This allows us to produce substantially improved sound results for phenomena such as rigid-body fracture.;Finally, we also consider sound from other, non-rigid phenomena; specifically, sound from physics-based animations of fire. We propose a hybrid sound synthesis algorithm combining physics-based and data-driven approaches. Our method produces plausible results for a variety of fire animations. Moreover, our use of data-driven synthesis grants users of our method a degree of artistic control.