Environmental sounds - such as the sound of two objects colliding - are affected by many physical factors, including the structure of the objects, their motion, and the surrounding environment in which sound echoes and reverberates. Humans can infer many of these physical factors from sound, but how they do so is unknown. Because multiple physical variables collectively determine any single sound, we consider hearing such sounds as a novel scene analysis problem. Presumably, humans must leverage the regularities of natural environments and objects to perform such inference. We measure environmental sound statistics by investigating the physical processes that generate them, and we test whether humans use these statistics to make causal inferences from natural sounds.