In order to create an immersive virtual world, it is crucial to incorporate a realistic aural experience that complements the visual sense. Physically-based sound simulation is a method to achieve this goal and automatically provides audio-visual correspondence. It simulates the physical process of sound: the pressure variations of a medium originated from some vibrating surface (sound synthesis), propagating as waves in space and reaching human ears (sound propagation). The perceived realism of simulated sounds depends on the accuracy of the computation methods and the computational resource available, and oftentimes it is not feasible to use the most accurate technique for all simulation targets. I propose techniques that model the general sense of sounds and their details separately and adaptively to balance the realism and computational costs of sound simulations. For synthesizing liquid sounds, I present a novel approach that generate sounds due to the vibration of resonating bubbles. My approach uses three levels of bubble modeling to control the trade-offs between quality and efficiency: statistical generation from liquid surface configuration,explicitly tracking of spherical bubbles, and decomposition of non-spherical bubbles to spherical harmonics. For synthesizing rigid-body contact sounds, I propose to improve the realism in two levels using example recordings: first, material parameters that preserve the inherent quality of the recorded material are estimated; then extra details from the example recording that are not fully captured by the material parameters are computed and added. For simulating sound propagation in large, complex scenes, I present a novel hybrid approach that couples numerical and geometric acoustic techniques. By decomposing the spatial domain of a scene and applying the more accurate and expensive numerical acoustic techniques only in limited regions, a user is able to allocate computation resources on where it matters most.