Object interactions – collisions, scraping and rolling – create many of the sounds that we hear in the world around us. These sounds are generated via lawful physical dynamics. Anecdotally, humans possess some intuitive knowledge of the physical generative processes underlying sound production, but little is known about the extent and nature of this knowledge. In this work, we study the human ability to make inferences about physical properties like mass, size, velocity and surface roughness from contact sounds, and test whether it is consistent with the use of a generative model of the underlying physics. Additionally, we propose a computational model that uses analysis-by-synthesis to infer object properties and their physical configurations using contact sounds as input, and compare it to human behavior.

We experience our soundscape in terms of physical events; for instance, a friend sweeping up after a plate crashed onthe floor. The underlying perceptual inferences are typically ill-posed: without constraints, there are infinite possiblecauses of the observed sound. Thus, a core task for cognitive science is specifying the variables we perceive along withthe constraints that allow them to be estimated. We identified sustained contact sounds (e.g., hands rubbing together,scraping a pan) as a rich domain with which to explore perceptual constraints. We developed a simple physics-basedsound-synthesis model that can generate a diverse set of realistic scraping sounds. We find that listeners perceive thegenerative physical variables from scraping sounds, including velocity, motion trajectory, and surface roughness. Furtherexperiments and acoustic analyses will address whether perception is constrained by a holistic generative model of soundor by invariant features that specify each perceived variable.

Everyday hearing requires inferring the causal factors that produce a sound, as when we separate the acoustic effectsof the environment (reverberation) from those of sound sources. Here we consider perceptual inferences from impactsounds, in which the resonance of a struck object provides cues to its material, but via acoustic effects that might benontrivial to disentangle from reverberation. We investigated whether and how humans separate the effects of objectresonance and reverberation in a material classification task. For comparison, we implemented a Bayesian observer thatinferred material from a generative model of object sounds without reverberation. Humans were robust to reverberation,whereas the model was not. However, human robustness was specific to reverberation consistent with the statistics ofnatural environments. The results suggest that humans use internal models of room and object acoustics to determine theirrespective contributions to sound, providing an example of causal inference in audition.