EEG hyperscanning data of 13 groups of three people engaging in different types of conversations

A hyperscanning dataset combining EEG, directional and omnidirectional audio from 13 groups of volunteers. Each group engaged in conversations elicited by three types of conversation sparkers. This dataset allows researchers to investigate, e.g., attentional dynamics unfolding during different types of social interactions to turn-taking, neurobiological markers of conversational engagement, etc.. ## Abstract This dataset contains preprocessed physiological data from a series of hyperscanning experiments conducted with 13 groups of three participants. The members of each group knew each other other prior to the study. Each participant wore a 32-channel EEG cap while enagaging in different types of conversations. Conversations were elicited by three types of conversation sparkers: consensus, free speech, and conflict. Each group took part in 12 individual conversations, which in the following are referred to as trials. After each trial, participants were asked to rate five different aspects of the conversations. The recordings contain the EEG signals sampled at 100 Hz as well as audio signals from the three directional microphones plus one omnidirectional microphone sampled at 44.1 kHz. They are paired with the behavioral ratings that were collected after each trial. ## Experimental Design ### Structure The recordings from each group of volunteers contain 12 trials organized across three conversation types: | Condition | Description | |-----------|-------------| | **Consensus** | Participants must reach group agreement on a topic | | **Free Speech** | Open, unstructured conversation | | **Conflict** | Participants discuss topics designed to elicit disagreement | After each trial, participants complete behavioral ratings. There are stored in `subj_rating.csv`. ### Behavioral engagement scores A table (`data/subj_rating.csv`) provides post-trial behavioral assessments, namely the ratings of each individual conversation by each individual participant. | Column | Description | |--------|-------------| | `Experiment` | Experiment identifier (Experiment 8–20) | Experiments 1-7 are not included. | `Subject` | Participant within the experiment (Subject 1, 2, or 3) | | `Trial` | Trial number (Trial 1–12) | | `Question` | Question prompt (Question 1–5) | | `Speech type` | Conversation condition: Free speech, Consensus, or Conflict | | `Rating` | Numeric post-trial rating score | ### Physical Setup - Subjects within each group knew each other prior to the session - EEG cap positioned on all participants; reference configuration via [CACS-32_REF] - Each participant was assigned a directional microphone positioned ~2 m away, with microphones spaced 120° apart around the group - Omnidirectional microphone: placed at the center of the group ## Data Acquisition | Modality | Channels / Mics | Sampling Rate | Format | |----------|-----------------|---------------|--------| | EEG | 96 channels (raw), split into 3 subjects × 32 channels each | 2,500 Hz → resampled to **100 Hz** | Continuous → trial-segmented | | Audio — Directional (mic1, mic2, mic3) | 1 channel per mic | **44.1 kHz** | Per-trial segments | | Audio — Omnidirectional (omni) | 1 channel | **44.1 kHz** | Per-trial segments | ### HDF5 File Contents Each experiment's HDF5 file contains bothe eeg and audio data: eeg_ica/ | Preprocessed EEG (ICA-cleaned, re-referenced) per trial {1–12} trial_1 … trial_12/ (12 groups) subject_1 (32, N) float64 {bad_channels, bad_windows} subject_2 (32, N) float64 {bad_channels, bad_windows} subject_3 (32, N) float64 {bad_channels, bad_windows}mic1/ trial1 … trial12 | Directional microphone 1 audio per trial {1–12} |mic2/ trial1 … trial12. | Directional microphone 2 audio per trial mic3/ trial1 … trial12. | Directional microphone 3 audio per trial omni/ trial1 … trial12. | Omnidirectional microphone audio per trial ### Processing Steps Embedded in HDF5 Files The EEG data stored under `eeg_ica/` has undergone the following preprocessing pipeline: 1. **Band-pass filtering**: High-pass at 0.5 Hz, low-pass at 20 Hz 2. **Resampling**: From raw 2,500 Hz down to 100 Hz 3. **Trial segmentation**: Continuous recording split into 12 trial windows (boundaries derived from event markers) 4. **Bad time-window detection**: Riemannian-distance outlier detector on overlapping 2-second windows (`BAD_riemann` annotations) 5. **Bad channel detection**: RANSAC-based noisy channel identification via `pyprep.NoisyChannels` 6. **Re-referencing & interpolation**: Average reference on good channels; bad channels interpolated 7. **ICA decomposition**: FastICA with automatic component count; non-brain/non-"other" ICLabel components rejected 8. **Channel interpolation**: Bad channels replaced via spherical spline interpolation ### Audio Data Audio data stored under each microphone group represents raw trial-level segments extracted from the continuous recordings. These are **not** preprocessed beyond extraction (no filtering, resampling, or artifact removal). ## Citation *Please cite this dataset when used in publications. DOI will be assigned upon Zenodo deposit.* P. Ochoa de Eribe Delgado, F. Marmel, R. Hannemann, T. Reichenbach (2026) EEG hyperscanning dataset: Neural mechanisms of speaking and comprehending during group conversations [Data set]. Zenodo. https://doi.org/[TBD]