Auditory perception of object properties as inverse acoustics

Abstract

Perception relies on regularities in sensory data that are caused by physical laws. In audition, physical laws governing object interactions constrain the structure of sounds. However, it remains unclear to what extent these constraints have been internalized by the brain to support auditory inferences about the world. To investigate whether physical constraints are modeled by the auditory system, we developed a Bayesian ideal observer for a physics-inspired generative model of impact sounds, and tested whether its judgments of the mass of a ball falling onto a board accorded with those of human listeners. To generate audio with the model, the time-varying impact force was convolved with measured impulse responses of the ball and board. The force varied parametrically with mass to simulate impacts involving balls of different masses. This relationship was fitted to measurements of a set of wooden balls with equal hardness. Inference was performed by Markov chain Monte Carlo sampling. The model accurately predicted most human judgments in a two alternative forced choice task using recorded audio. However, the model underestimated mass when the ball material was harder than wood. The results suggest that additional physical parameters (e.g., hardness) must be modeled to account for human perception.