SoundBubble‑EEG: A Multi‑Scenario Dataset for Naturalistic Multi‑Talker Auditory Attention Decoding

SoundBubble‑EEG: A Multi‑Scenario Dataset for Naturalistic Multi‑Talker Auditory Attention Decoding Overview This dataset contains 128-channel EEG recordings from 30 participants performing a selective auditory attention task in three ecological sound bubble scenarios: Restaurant, TV, and Meeting. Participants were instructed to attend to a spatially cued audio stream (left or right) presented via in-ear headphones while ignoring the competing stream from the opposite direction. Dataset Structure 30 participants (sub-01 to sub-30) 3 runs per participant, each corresponding to one acoustic scenario: Run 01: Restaurant Run 02: TV Run 03: Meeting Note: In the raw EEG files (events.tsv and preprocessed .fif files), scenario labels use the original acquisition names. The mapping to the canonical scenario names used in this dataset is as follows: "Cafe" → Restaurant "Office" → Meeting "TV" → TV (unchanged) EEG: 128-channel HydroCel GSN 128 cap, recorded at 250 Hz with an EGI NA-400 amplifier Audio stimuli: delivered via SONY MDR-EX15LP in-ear headphones (binaural spatial audio) Task In each trial, a visual cue indicated the target direction (left or right). Participants attended to the sound source from that direction and ignored the competing source. The task was designed to simulate real-world "sound bubble" listening conditions. Participants 30 healthy adults with self-reported normal hearing (20 male, 10 female; age range 18–30 years). Handedness: 28 right-handed, 2 left-handed. Equipment EEG amplifier: EGI NA-400 EEG cap: HydroCel GSN 128 1.0 (128 channels) Stimulus delivery: SONY MDR-EX15LP in-ear headphones Software: Net Station 5.2, E-Prime 2.0 Event Timestamp Alignment Due to a bug in the E-Prime trigger configuration during early data collection, the TBEG event markers embedded in the EGI MFF files for sub-01 through sub-09 were not correctly aligned with actual stimulus onset times. To recover usable data, a dual-source alignment procedure was applied to these nine subjects during preprocessing. Problem: The E-Prime software failed to send correctly timed trigger pulses to the EEG amplifier, causing the TBEG markers recorded in the MFF XML to be misaligned with true stimulus onsets. Solution — Dual-Source Linear Regression Alignment: Two independent time sources were cross-referenced to reconstruct accurate stimulus onset samples: E-Prime behavioral log (SoundOut1.OnsetTime): millisecond-precision stimulus onset times recorded by the stimulus presentation software. EGI hardware acquisition log (TBEG entries): timestamps written by the EEG amplifier's own clock at trial boundaries. A histogram of pairwise time differences between the two sources was used to identify the dominant clock offset, after which matched pairs were fitted with a linear regression model. This model was then applied to all trials to predict their corresponding EEG sample indices. Validation — Leave-One-Subject-Out Cross-Validation (LOSO-CV): The accuracy of the dual-source alignment was rigorously validated using LOSO-CV on sub-10 through sub-30 (21 subjects, 525 trials), where the MFF XML TBEG markers serve as ground truth. For each held-out subject, a run-level clock offset correction was learned exclusively from the remaining 20 subjects and applied to the held-out subject's predictions — ensuring strict separation between training and test data with no circular reasoning. Results: after LOSO-CV correction, the mean timing error was +0.69 ms (effectively zero bias), MAE 22.4 ms, and Std 27.3 ms. The residual jitter (Std ≈ 27 ms) represents the physical timing noise floor of the recording hardware (OS scheduling latency, E-Prime clock resolution, and EGI log write delay) and cannot be reduced by post-processing. Practical sufficiency for AAD research: All trials in this dataset use 120-second continuous epochs. A timing uncertainty of 22 ms corresponds to less than 0.02% of the epoch duration and is negligible for correlation-based auditory attention decoders, which are insensitive to sub-100 ms onset jitter at this epoch length. For sub-10 through sub-30, the E-Prime trigger logic was corrected prior to recording. Event onsets for these subjects were extracted directly from the MFF XML without any alignment procedure. Practical note for users: The onset column in each _events.tsv file already reflects the corrected, alignment-adjusted timestamps for all subjects. No additional correction is needed when working with the BIDS events files. License This dataset is released under the CC0 1.0 Universal license. Citation Please cite the associated paper (in preparation) when using this dataset.