AISKI speech stimuli: AI-generated Korean and English two-talker speech for auditory attention decoding

Licence: Custom licence — free download and non-commercial use; the stimuli may not be used to train text-to-speech, speech generation, voice cloning or similar generative speech models (see LICENSE file in this record; terms follow the Typecast TTS service of NeoSapience, Inc.) Related record: Dataset - Preprocessed EEG, behavioral responses, participant information and benchmark splits — https://doi.org/10.5281/zenodo.23589425 Overview This record contains the two-talker speech stimuli, and their scripts, used in AISKI, a multi-site bilingual EEG dataset for auditory attention decoding (AAD). The stimuli were produced with generative AI: short narrative stories and a matched comprehension question for each story were generated with a large language model (GPT-4.1) in Korean and translated into English. Each story was synthesized with the Typecast text-to-speech service in both languages with the same voice, using a pool of 137 voices. The EEG recordings, behavioral responses and benchmark split assignments collected with these stimuli are distributed in a separate Zenodo record (see "Related work" below). Contents Two archives with identical structure: Korean_AISKI.zip and English_AISKI.zip. Korean_AISKI/ (or English_AISKI/) day1block1/ ... day3block3/ (9 block folders) target/ day{d}block{b}_stimulus{NN}.wav (NN = 01–30) interference/ day{d}block{b}_stimulus{NN}.wav mixture/ day{d}block{b}_stimulus{NN}.wav mixture_info.csv mixture/ — the two-talker mixture that was actually presented to participants (target + interference) target/ — the target talker's stream alone interference/ — the interfering talker's stream alone mixture_info.csv — per-stimulus metadata for that block, columns: stimulus_num, name_target (target voice), name_interference (interfering voice), utterance (script spoken by the target talker), question (comprehension question), option 1–option 4 (answer options), answer_index (number of the correct option), answer (text of the correct option) All in all: 9 blocks × 30 stimuli = 270 stimuli per language (135 unique voice pairs). Stimulus details Audio format: WAV, 44.1 kHz, mono. Duration and structure: every file is 34 s. The first 4 s is a pre-cue in which the target talker says "Please focus on this voice", followed by 30 s of speech. Role switching: within each block, stimuli 1–15 and 16–30 use the same voice pairs with the target and interference roles swapped, i.e. the mixture is the same but the talker to be attended changes. Each script therefore occurs once as the target and once as the interfering stream. Bilingual correspondence: day{d}block{b}_stimulus{NN} in Korean_AISKI and English_AISKI share the same voices. Use in the paper: all nine blocks of Korean_AISKI were presented to the Korean-native participants (SNU, KIST, UI), and day1block1 and day1block2 of English_AISKI were presented to the English-native participants at UI. The remaining English blocks were not used for EEG recording but are released for completeness. Linking to the EEG record In the EEG record, epoch_events{i} of a session's *_eeg.mat file gives the stimulus number NN of trial i. The session number (ses-{SS} in the file name) gives d, and the idxBlock column of row i in the corresponding *_task-ASA_beh.csv file gives b; the audio presented on that trial is then day{d}block{b}_stimulus{NN}.wav in this record.