Data and code for "Acoustic regularities define perceptual and cortical representations of voice-likeness" (Hect et al., Current Biology, 2026)

Description: This repository contains the analysis code, preprocessed neural data, and all intermediate results needed to reproduce the figures and statistical analyses reported in Hecht et al. (2026), Current Biology. Overview of the study. Humans readily recognize voices across diverse and noisy environments, yet the neural basis of voice perception remains debated. This study tested whether voice perception is organized along a continuous, category-defining acoustic dimension and whether this structure is hierarchically encoded across the human auditory ventral stream. Using intracranial electroencephalography (iEEG/sEEG) recorded from 39 participants performing auditory n-back tasks, combined with perceptual ratings from 253 online listeners, we show that a linear discriminant axis derived from 90 acoustic features of natural sounds predicts graded voice-likeness ratings for both natural and novel synthetic stimuli. Neural population responses across six auditory ventral stream regions (primary auditory cortex, belt, parabelt, superior temporal gyrus, ventrolateral prefrontal cortex, and orbitofrontal cortex) mirror this acoustic-perceptual organization, with graded structure emerging hierarchically from primary to association cortex and generalizing to synthetic stimuli without true category membership. Contents. Code. The full MATLAB analysis pipeline is available at https://github.com/pbe-lab/Codeshare_Hect2026.git. The main entry point is main_analysis.m, which includes a checkpoint system that saves and loads results from expensive computations (permutation tests, bootstrap resampling, time-varying LDA) so that figures can be regenerated without rerunning the full pipeline. Data files. The following preprocessed data files are included in this upload: File Contents hecht2026_behavioral.mat Perceptual voice-likeness ratings from 253 online listeners (Gorilla/Prolific), stimulus sort indices, quartile group labels, and YAMNet DNN category predictions for all 394 stimuli hecht2026_acoustic.mat 90-dimensional acoustic feature matrices (88 GeMAPS features + temporal and frequency correlation decay coefficients) for natural and synthetic stimuli, acoustic LDA model weights and projections, and permutation test results hecht2026_neural_lfp.mat Preprocessed local field potential (LFP) data matrices for both tasks: baseline-normalized, repeat-averaged, downsampled to 200 Hz, and sorted by voice-likeness rating. Dimensions: [nTimepoints x nChannels x nStimuli]. Includes channel metadata (HCPex parcel labels, hemisphere, patient-channel identifiers in MNI space) hecht2026_results_roi.mat All ROI-level analysis results: LDA cross-validated accuracy and permutation statistics, Spearman rank correlations between neural axis projections and voice-likeness ratings (natural sounds, voice-only, nonvoice-only, and synthetic sounds), bootstrap 95% confidence intervals, FDR-corrected q-values, between-ROI comparison p-values, and LME model comparison statistics (AIC, BIC, likelihood ratio tests) hecht2026_results_singlechan.mat Per-channel LDA accuracy, rating correlations, acoustic-neural alignment correlations, and model comparison statistics for all electrodes in auditory ventral stream regions, with MNI surface coordinates for brain map generation hecht2026_results_timedomain.mat Sliding-window (150 ms, 100 ms step) LDA classification accuracy and time-resolved Spearman rank correlations for all ROIs and both tasks, along with time-varying single-channel model comparison results hecht2026_results_acneural.mat ROI-level and single-channel correlations between acoustic LDA axis projections and neural LDA axis projections for synthetic sounds, with permutation-derived p-values hecht2026_lme.mat Long-format MATLAB tables and fitted linear mixed-effects model coefficients for the analysis predicting LFP amplitude from continuous voice-likeness ratings with random intercepts for participant Stimuli. The 250 synthetic sound textures generated for this study are included as audio files (.wav, 44.1 kHz). The 144 natural sounds from the Belin voice localizer corpus are not redistributed here; they are available from the original authors (Belin et al., 2000, Nature; https://doi.org/10.1038/35002078). Raw iEEG recordings. Raw neural recordings are not included in this deposit due to IRB data sharing restrictions (UPMC IRB protocol STUDY20030060). The preprocessed data matrices in hecht2026_neural_lfp.mat are derived from these recordings and contain no protected health information. Reproducing the analyses. Download codebase from Github. git clone https://github.com/pbe-lab/Codeshare_Hect2026.git To regenerate all figures from saved results without rerunning computations, download the data files and run the visualization sections of main_analysis.m (after loading the relevant .mat files. To rerun analyses from the preprocessed neural data, load hecht2026_neural_lfp.mat and run Sections B onward in main_analysis.m. Full instructions are included in README.md and in the hecht2026_metadata.mat loading instructions field. Software requirements. MATLAB R2024b with the Statistics and Machine Learning Toolbox, Bioinformatics Toolbox, and Audio Toolbox. Python 3.10 with openSMILE >= 3.0 is required only to re-extract acoustic features from raw audio. Related publication. Hecht JL, Rupp KM, Ghuman AS, Holt LL, Abel TJ (2026). Acoustic regularities define perceptual and cortical representations of voice-likeness. Current Biology. https://doi.org/[to be assigned]. Keywords: voice perception, intracranial EEG, auditory cortex, voice-likeness, acoustic features, linear discriminant analysis, hierarchical processing, sEEG, human electrophysiology, synthetic sounds, psychoacoustics