neukym/kymata-soto

Dataset Card for the Kymata SOTO Dataset

This dataset is comprised of magnetoencephalography and electroencephalography recordings of English and Russian speakers listening to podcasts of their native languages.

Details of the dataset can be found here: https://doi.org/10.1038/s41597-026-06579-8

Dataset Details

Dataset Description

The Kymata Soto Language Dataset comprises raw electroencephalographic (EEG) and magnetoencephalographic (MEG) recordings from 15 native Russian speakers and 20 native English speakers as they listened to approximately seven minutes of conversational speech in their respective native languages[cite: 9]. [cite_start]Each participant heard the same conversational speech stimulus multiple times[cite: 10]. [cite_start]The dataset includes transcriptions of the recordings, along with timestamp annotations for each phoneme and word[cite: 11]. [cite_start]Organized according to the Brain Imaging Data Structure (BIDS), this dataset facilitates in-depth research into brain responses to naturalistic speech.

• Curated by: Chen Tianyi Yang, Oliver Parish, Anastasia Klimovich-Gray, Cai Wingfield, William D. Marslen-Wilson, Chao Zhang, Alexandra Woolgar, and Andrew Thwaites.
• Funded by: UKRI MRC (SUAG/093/G116768), ERC Advanced Grant (Neurolex), and the National Natural Science Foundation of China (NSFC) (62476151).
• Shared by: The dataset authors, with corresponding author Andrew Thwaites.
• Language(s) (NLP): English and Russian.
• License: Creative Commons Attribution 4.0 International Licence (CC-BY).

Dataset Sources

• Original Repository: OSF (https://osf.io/qdzgh/).
• Paper: https://doi.org/10.1038/s41597-026-06579-8.

Uses

Direct Use

The dataset is particularly suited for investigating intra-language linguistic comparisons and cross-language differences in speech perception. It enables the examination of complex language processing mechanisms and provides data for applications such as decoding speech from brain recordings.

Out-of-Scope Use

Direct cortical localisation mapping using the Russian dataset's native structural MRI files is out of scope, as these files are not shared; researchers will instead need to use an average cortical mesh like Freesurfer's FSaverage. Additionally, as the visual stimulus was a dynamic dot field designed to avoid eliciting visuosemantic processes, it is not suited for complex visual-semantic mapping.

Dataset Structure

The dataset is organized according to the Brain Imaging Data Structure (BIDS) version 1.9.0.

• The root directory includes a dataset_description.json file and MEG-machine-specific calibration files.
• A stimuli folder contains the original audio podcasts and visual stimuli frame-by-frame.
• Individual participant data is housed in sub-XXX folders, where 'R' denotes native Russian speakers and 'E' denotes native English speakers.
• Participant folders contain the raw MEG data and, for the English dataset, anatomical MRI data.

Dataset Creation

Curation Rationale

This dataset was created to address a shift in language comprehension research toward more naturalistic contexts. Instead of analyzing responses to isolated words or scripted narrations, this dataset focuses on continuous, spontaneous conversational speech, which exhibits more variable pacing and is crucial for understanding real-world language processing.

Source Data

Data Collection and Processing

• Concurrent EEG and MEG data were recorded while participants listened to podcasts through earpieces, sampled at 1,000 Hz.
• MEG data were recorded using a 306-channel Elekta-Neuromag Vector View or MEGIN Truix system in a magnetically shielded room.
• EEG data were recorded from 70 Ag-AgCl electrodes placed in an elastic cap.
• The MEG dataset is shared raw (not preprocessed); users will need to pass the data through Signal-Space Separation (SSS) and/or maxwell filtering to remove environmental artefacts.

Who are the source data producers?

The dataset features brain recordings from 15 Russian-speaking adults (7 male; 8 female, aged 18–30) and 20 English-speaking adults (11 male; 9 female, aged 18–35) recruited from the Cambridge University Medical Research Council Cognition and Brain Sciences Unit. The auditory stimuli are edited excerpts from BBC radio-podcast studio discussions (You're Dead to Me and Биби Сева).

Annotations

Annotation process

Transcriptions for both recordings were generated using the Whisper large v2 model with default parameters, and were subsequently corrected manually. Timestamp information at both the word and phoneme level was obtained by feeding these transcriptions into the Montreal Forced Aligner.

The triggers for each repetition of the audio stimulus is '3', found in ST101.

Who are the annotators?

The transcriptions were generated via automated tools and manually corrected by the authors of the study.

Personal and Sensitive Information

Efforts were made to protect participant anonymity. The individual structural MRI data of the English dataset is provided only after being defaced using PyDeface. The MRI structural data for the Russian dataset is not shared at all due to the specific wording of the participant consent forms.

Citation

BibTeX:

@article{Yang2026, author = {Yang, ChenTianyi and Parish, Oliver and Klimovich-Gray, Anastasia and Wingfield, Cai and Marslen-Wilson, William D. and Zhang, Chao and Woolgar, Alexandra and Thwaites, Andrew}, title = {Kymata Soto Language Dataset: an electro-magnetoencephalographic dataset for natural speech processing}, journal = {Scientific Data}, year = {2026}, volume = {13}, number = {1}, pages = {254}, issn = {2052-4463}, doi = {10.1038/s41597-026-06579-8}, url = {https://doi.org/10.1038/s41597-026-06579-8} }

APA:

Yang, C., Parish, O., Klimovich-Gray, A. et al. Kymata Soto Language Dataset: an electro-magnetoencephalographic dataset for natural speech processing. Sci Data 13, 254 (2026). https://doi.org/10.1038/s41597-026-06579-8