Distinguishing fine structure and summary representation of sound textures from neural activity

Abstract

The auditory system relies on local and global representations, depending on the amount of entering temporal details. When local information exceeds system constraints, acoustic details are summarized into a set of average statistics and global structure emerges. Whether computations subtending local and global representations of sounds can be distinguished at the cortical level is unknown. A computational auditory model was employed to extract auditory statistics from natural sound textures, such as fire, wind, and rain, and generate synthetic exemplars in which we had control of local and global properties. Participants were passively exposed to auditory streams with triplets of sounds including two standards and an oddball while the EEG was recorded. In Local Discrimination (exp 1) the oddball sound varied from standards only for its local features; in Global Discrimination (exp 2), it varied for its global structure. Different sound durations were employed to manipulate the amount of acoustic information and statistical variability between standards and oddballs. We expected short and long sounds to engage local (exp 1) or global (exp 2) representations, respectively. As predicted, neural responses associated with local discriminations exceeded global ones for short sounds, whereas the opposite pattern emerged for longer sounds. Electrical source estimates revealed that local discriminations mainly engaged auditory regions of the right hemisphere, while global discriminations of the left hemisphere. These findings revealed distinct neural correlates of acoustic discriminations guided by local or global structures. Ultimately results demonstrated that the auditory system developed a functional architecture to automatically detect changes occurring in the environment based on these two modes of sound representation.SIGNIFICANT STATEMENTPrior to this study, it was not known whether it was possible to directly measure auditory cortical representations of sound discriminations depending on local or global acoustic properties. Results unveiled the existence of a mechanisms which pre-attentively determines at which timescale (local or global) a change in a stream occurred based on statistical variability between sound excerpts. By performing source reconstruction, we additionally revealed that the response to local change detection pertains to different cortical structures as compared to global one, displaying also hemispheric asymmetries. These findings expand our knowledge on fundamental mechanisms characterizing auditory system functioning.

Hemispheric asymmetries in auditory cortex reflect discriminative responses to temporal details or summary statistics of stationary sounds

“…The two effects emerged over similar, but not identical, scalp locations, possibly indicating different sources. Moreover, different oscillatory profiles were associated with local or summary changes, with local features being encoded by faster oscillations than summary statistics (Berto et al, 2022). Interestingly, these results emerged without an explicit task.…”

“…The functional role of local features and summary statistics in the processing of stationary sounds has been evaluated in humans through computational synthesis approaches (Berto et al, 2021(Berto et al, , 2022McDermott & Simoncelli, 2011;McWalter & McDermott, 2019;Norman-Haignere & McDermott, 2018;Zuk et al, 2020); for instance, it was possible to disentangle the processing of local features from summary statistics by creating synthetic sounds with the same auditory statistics but different acoustic details (Berto et al, 2021(Berto et al, , 2022McDermott et al, 2013).…”

“…Specifically, a previous study employed this approach to isolate neural signatures of local features and summary statistics of sound textures (Berto et al, 2022). In this study, brain activity was recorded using electroencephalography (EEG) while participants were exposed to streams of sounds comprising changes based on local features or summary statistics and the duration of the sound excerpt could be short, medium, or long to manipulate the reliability of local features or summary statistics processing.…”

“…In the current study, we leveraged the high spatial resolution of the magnetoencephalography (MEG) to provide reliable source estimates of the automatic processing of local and summary acoustic information, without losing temporal precision. We implemented a validated protocol (see Berto et al, 2022) and used short (40ms) and long (478ms) sounds presented in separate continuous streams at a constant rate (one sound every half a second). Within each stream, sounds appeared as a series of triplets, with the first two being identical and the third varying for its local features or summary statistics, depending on the experimental context.…”

Brain and grammar: revealing electrophysiological basic structures with competing statistical models

“…First, EEG data were downsampled at 250 Hz to reduce computational time. Then, 2 pre-processing steps were carried out, using a semiautomatic pipeline [12], [13], [14]. The pre-processing was divided into two steps to minimize the removal of brain activity while maximizing the quality of the Independent Component Analysis (ICA) decomposition.…”