Cochlea to categories: The spatiotemporal dynamics of semantic auditory representations

Cognitive Neuropsychology, 2022

Click for Full Article PDF and Journal Link

Matthew X. Lowea,b,* Yalda Mohsenzadeha,c,d,e,* Benjamin Lahnera Ian Charestf,g Aude Olivaa, Santani Tenga,h,
aComputer Science and Artificial Intelligence Lab (CSAIL), MIT, Cambridge, MA, USA
bUnlimited Sciences, Colorado Springs, CO, USA
cThe Brain and Mind Institute, The University of Western Ontario, London, Canada
dDepartment of Computer Science, The University of Western Ontario, London, Canada
eVector Institute for ArtificialIntelligence, Toronto, Ontario, Canada
fDépartement de Psychologie, Université de Montréal, Montréal, Canada
gCenter for Human Brain Health, University of Birmingham, Birmingham, UK
hSmith-Kettlewell Eye Research Institute (SKERI), San Francisco, CA, USA
* denotes equal first author contribution
† denotes equal senior author contribution
barking dog teaser
Figure 1: Big Picture - When and where does the brain transform sound waves into something we can understand?

Cochlea to categories: condensed version

Introduction: Our auditory system transforms sounds from a mixture of frequencies (e.g. high and low tones) into a meaningful concept (e.g. a barking dog) (Figure 1). But when does this transition occur and which brain regions are involved? In this work we address when and where acoustically-dominated representations change to semantically-dominated representations in the human brain. Our multimodal imaging approach reveals distributed and hierarchical aspects of this process.

Methods: We present 80 different natural sounds across 4 semantic categories (voices, animals, objects, scenes) to 16 participants as we collect temporally-resolved MEG and spatially-resolved fMRI brain responses. Using the MEG-fMRI Fusion method, we achieve high spatiotemporal resolution to determine when and where the sounds are represented throughout the brain (Figure 2B, also see Supplementary Movie 1). We compare our experimentally-collected representations against idealized models of a purely acoustic representation (based on cochleagram differences) and a purely semantic representation (based on semantic category membership) to measure which regions show preference to acoustic features or semantic features (Figure 3A).

Results: We find that peak latencies in primary and non-primary auditory regions (PAC, TE1.2, PT, and PP) were temporally indistinguishable at ~115ms, voice-selective (TVAx and LIFG) and MPA were temporally indistinguishable significantly later at ~200ms, and high-level visual regions (FFA, LOC, PPA) were temporally indistinguishable later still at ~300ms (Figure 2AB). The sounds' acoustic features are most strongly coded in primary auditory regions early in time, and the sounds' semantic category features are most strongly coded in secondary auditory regions and even high-level visual regions later in time (Figure 3AB). High-level auditory and visual regions additionally show preference to certain sound categories.

Conclusion: Our findings show that across the auditory cortex (and especially in extra-auditory regions), the temporal progression of brain responses strongly correspond to a region's spatial distance from PAC and increasingly acoustic-to-semantic representations. At a finer level within parts of the auditory cortex (especially the primary and non-primary auditory ROIs), numerous regions simultaneously respond to bring about the acoustic-to-semantic transformation. Thus, the human auditory cortex resembles both a hierarchical processing stream and a distributed processing stream dissociable by space, time, and content.
ROI Fusion
Figure 2: MEG-fMRI ROI Fusion (A) The location of 11 ROIs in the whole brain. (B) The similarity curves between the MEG Representational Dissimilarity Matrices (RDMs) at each millisecond and the fMRI RDMs for each ROI show a spatiotemporal view of auditory processing. PAC, TE1.2, PT, and PP show a similar peak latency (~115ms), followed by TVAx, LIFG, and MPA (~200ms), and lastly FFA, PPA, and LOC (~300ms). EVC shows no significance, as expected.
ROI Fusion
Figure 3: Semantic dominance over space and time (A) The correlation between the MEG RDMs at each millisecond and the idealized cochleagram RDM and category RDM shows how semantic dominance increases over time. (B) The correlation between the fMRI RDMs for each ROI and the idealized cochleagram RDM and category RDM shows semantic dominance increases in non-primary auditory ROIs.

MEG-fMRI Fusion Movie

Supplementary Movie 1: MEG-fMRI Fusion Movie - We correlate the MEG RDMs (of size 80 stimuli x 80 stimuli) at each millisecond with the fMRI RDMs (of size 80 stimuli x 80 stimuli) at each voxel to measure when and where the auditory representations are most similar. In this way we obtain a movie that depicts the "correlational flow" of human auditory processing.


The study was conducted at the Athinoula A. Martinos Imaging Center, MIBR, MIT. We thank Dimitrios Pantazis for helpful discussion, and Michele Winter for help with stimulus selection and processing.
This work was funded by an Office of Naval Research Vannevar Bush Faculty Fellowship to A.O. [grant number N00014-16-1 3116]. S.T. was supported by an NIH training grant to SmithKettlewell Institute [grant number T32EY025201] and by the National Institute on Disability, Independent Living, and Rehabilitation Research [grant number 90RE5024-01-00].
No potential conflict of interest was reported by the author(s).

Data Availability

We are actively working on releasing the MEG and fMRI data here soon! Thank you for your patience.