AISKI: A multi-site, bilingual EEG dataset for auditory attention decoding

Overview AISKI is a multi-site, bilingual EEG dataset for auditory attention decoding (AAD). Twenty adults listened to two simultaneous talkers and attended to one of them while EEG was recorded at three laboratories: Seoul National University (SNU, South Korea), the Korea Institute of Science and Technology (KIST, South Korea) and the University of Iowa (UI, USA). The speech stimuli were produced with generative AI (LLM-written stories and comprehension questions, rendered with text-to-speech in Korean and English using 137 voices whose identity is shared across the two languages), so that the dataset supports generalization tests across unseen speakers, subjects, recording sites and language. This record contains the preprocessed EEG, trial-wise behavioral responses, participant information and the train/validation/test split assignments used in the accompanying paper. The speech stimuli (audio) and their scripts are distributed in a separate Zenodo record under a different licence (see "Related works" below). Experimental design (summary) Participants (n = 20): SNU, 6 Korean-native (Korean stimuli); KIST, 6 Korean-native (Korean stimuli); UI, 3 Korean-native residing in the USA (Korean stimuli) and 5 English-native (English stimuli). Task: diotic two-talker listening. Each trial is 34 s of audio: a 4-s pre-cue in the target speaker's voice ("Please focus on this voice") followed by a 30-s mixture of a target and an interfering talker. The two talkers share the same spatial location and the target can be male or female, so the target is identified by voice and content only. Behavioral responses per trial: (1) attention rating — "How well did you focus on the target speaker?" on a 7-point scale (1 = Not at all, 7 = Very well) (2) a four-alternative comprehension question about the target speech Sessions: Korean cohorts (SNU, KIST, UI-Korean): 3 sessions (days) × 3 blocks × 30 trials = 270 trials per participant. All Korean-native participants heard the same 270 mixtures; the assignment of mixtures to the nine blocks was fixed, and the trial order within each block was randomized per participant. English cohort (UI-English): 1 session × 2 blocks × 30 trials = 60 trials, using the English versions of the day1block1 and day1block2 stimuli. EEG systems: SNU and UI — BioSemi ActiveTwo; KIST — Tucker-Davis Technologies system. Total: about 36.25 h of EEG paired with attended speech. Preprocessing pipeline 60-Hz line-noise removal, downsampling (128 Hz), 0.1-Hz high-pass filter, common-average re-referencing, and epoching from 0 to 34 s relative to pre-cue of the speech stimulus onset (4-s pre-cue + 30-s mixture) No ICA and no trial rejection were applied In addition to the full-montage version, a 56-channel version retaining only the scalp positions common to all three sites (channels reordered to match across sites) is provided for cross-site analyses. Contents Four archives, one per site–language dataset: KOR_SNU_AISKI.zip, KOR_KIST_AISKI.zip, KOR_IOWA_AISKI.zip, ENG_IOWA_AISKI.zip. Each archive contains: {LANG}_{SITE}_behav/ — one CSV per participant and session: sub-{LANG}{SITE}{NNN}_ses-{SS}_task-ASA_beh.csv (e.g. sub-ENGIOWA001_ses-01_task-ASA_beh.csv) {LANG}_{SITE}_preproc_{N}ch/ — one MAT file per participant and session: sub-{LANG}{SITE}{NNN}_ses-{SS}_task-ASA_desc-preproc{N}_eeg.mat (e.g. sub-ENGIOWA001_ses-01_task-ASA_desc-preproc64_eeg.mat), where {N} is the number of EEG channels in that version (full montage or the shared 56-channel montage) Additional files: participants.tsv / participants.json — participant ID, institution, stimulus language, nationality, country of residence, sex, age, session dates and audiometry result Unseen_Speaker_split_assignments_KOR_SNU.csv — trial-level partition of the 270 SNU stimuli (development / seen-speaker test / unseen-speaker test) with five-fold CV assignment Unseen_Speaker_devset_238trials_KOR_SNU.xlsx — per-block summary of the same partition Unseen_Subject_split_assignments_KOR_SNU.xlsx — leave-one-subject-out assignment used for the unseen-subject benchmark File formats Behavioral CSV (*_task-ASA_beh.csv), one row per trial in presentation order: idxBlock (block number within the session) idxTrial (trial number within the block) question (comprehension question shown; Korean for the Korean cohorts, English for the English cohort) attentionRating (1–7) selectedAnswer (option chosen, 1–4) correctAnswer (correct option, 1–4) hit (1 = correct, 0 = incorrect) EEG MAT (*_eeg.mat), epoched and preprocessed EEG for one session: eeg_data — trials × channels × time (e.g., time = 4352 samples = 34 s × srate) 60 trials per file for UI-English (2 blocks × 30), 90 trials per file for the Korean cohorts (3 blocks × 30) channel_names — channel labels srate — sampling rate (128 Hz) times — time vector in seconds, 0 to 33.99 s relative to pre-cue onset epoch_events — the stimulus number presented on each trial Linking EEG, behavior and stimuli. Trials are stored in presentation order. Trial i of eeg_data corresponds to row i of the behavioral CSV of the same session (i.e. idxBlock and idxTrial), and epoch_events{i} is the stimulus number NN of the audio file day{d}block{b}_stimulusNN.wav in the stimulus record (see "Related works" below), where d is the session number and b is idxBlock. The target/interference speakers, script, question and answer options for each stimulus are listed in the mixture_info.csv of the corresponding day{d}block{b} folder of the stimulus record. Split assignment files (SNU dataset, used for the benchmarks in the paper). Unseen_Speaker_split_assignments_KOR_SNU.csv has one row per trial (6 subjects × 270 trials; subject 1–6 = sub-KORSNU001–006) with subject, day, block, trial, stimulus number, attended and interference speaker names, a pair_key for the competing-voice pair, fixed_partition (dev: 238 trials; test_seen_speaker: 16 trials whose voices also appear in dev; test_unseen_speaker: 16 trials whose voices appear in no dev or seen-speaker trial), cv_fold (five stimulus-level folds, 0–4, within dev, identical across participants) and the resulting train/validation/test role of each trial in each of the five folds. Because the partition is defined at the stimulus level and all Korean-native participants heard the same stimuli, it applies to the KIST and UI-Korean datasets as well. Unseen_Speaker_devset_238trials_KOR_SNU.xlsx summarizes, per block, how many of the 15 voice pairs fall into each fold and test set. Unseen_Subject_split_assignments_KOR_SNU.xlsx lists, for each held-out test subject, the five inner folds (validation subject / training subjects) of the nested leave-one-subject-out procedure.