Brain Representations of Natural Sound Statistics

Abstract

Natural sound textures (e.g. rain, crackling fire) are perceptually defined by time-averaged summary statistics. While previous studies have examined neural responses to natural sounds, little is known regarding the neural processing of the statistics underlying these sounds. To study neuronal correlates of these statistics, we measured brain responses to synthetic sound textures in which statistical structure was systematically varied while preserving the texture category. Using two fMRI experiments (males and females), we examined neural responses along the ascending auditory pathway, within auditory cortex and medial temporal lobe (MTL) regions implicated in pattern analysis. In Experiment 1, we parametrically varied the full set of texture statistics, creating sounds with different levels of naturalness. In Experiment 2, we selectively manipulated high-level statistics (cochlear skewness and kurtosis, cochlear and modulation correlations) while holding low-level statistics (cochlear mean and modulation power) constant. Increasing texture naturalness produced graded increases in BOLD responses across bilateral primary and nonprimary auditory cortex in both experiments, although overall responses were weaker in Experiment 2. This reduction suggests that low-level statistics contribute substantially to response magnitude, although higher-order statistics are sufficient to elicit graded responses. We also observed modulation in MTL regions, including entorhinal cortex, in Experiment 1. Moreover, functional connectivity between hippocampus and auditory cortex increased for more degraded (less natural) textures, suggesting a modulatory rather than representational role for MTL in texture processing. Together, these findings show that sensitivity to texture statistics is distributed across the auditory cortex and highlight MTL-auditory interactions when texture structure is ambiguous.