DECOMEG — Brain Activity During Typing (MEG & EEG)

Non-invasive brain recordings (magnetoencephalography, MEG; and electroencephalography, EEG) of healthy adults typing briefly-memorized sentences on a QWERTY keyboard. This is the dataset underlying Brain2Qwerty (Lévy et al., 2025) and its companion neuroscience study (Zhang et al., 2025).

Summary

• Participants: 35 healthy adult volunteers recruited at the Basque Center on Cognition, Brain and Language (BCBL), San Sebastián, Spain. All native Spanish speakers, right-handed, and skilled typists (selected for typing accuracy ≥ 80%). Cohort: 23% men / 77% women, mean age 31.6 ± 5.2 years. Five participants took part in both EEG and MEG sessions.
• Task: Each trial had three phases — read → wait → type. A Spanish sentence was shown word-by-word (rapid serial visual presentation, RSVP); after the last word a fixation cross appeared for 1.5 s; its disappearance cued the participant to type the sentence from memory without any on-screen feedback. Each session used 128 unique declarative Spanish sentences of 5–8 words.
• Languages / stimuli: Spanish sentences. MEG: ~5.1K sentences / ~193K characters. EEG: ~4K sentences / ~146K characters.
• Keyboard: A custom MR-compatible QWERTY keyboard (HybridMojo LLC) with non-ferromagnetic silver-spring key mechanisms, to avoid magnetic artifacts in the MEG.

Recording devices

Per-participant recording time: EEG 0.88 ± 0.02 h, MEG 0.93 ± 0.01 h (≈17.7 h EEG and ≈21.5 h MEG of typing in total).

Directory structure

pinet2024_public/
├── MEG/
│   ├── FIF/      # raw continuous MEG (Elekta/Megin .fif), one directory per recording;
│   │             #   each holds the typing blocks (block1.fif, block2.fif) + a tapping localizer
│   └── logs/     # behavioral logs (MATLAB .mat): stimuli, keystrokes, and timing
└── EEG/
    ├── EEG/      # raw EEG in BrainVision format (.eeg / .vhdr / .vmrk)
    └── logs/     # behavioral logs (MATLAB .mat): stimuli, keystrokes, and timing

File counts (this release)

Repeated participants (MEG)

Some people took part in more than one MEG session and appear under different subject IDs. The following IDs belong to the same person:

Merging these (and excluding S23, who had a metallic implant) yields 19 unique MEG participants. The Brain2Qwerty V1 pipeline applies exactly this mapping in its SpanishBCBLPreprocessing event transform.

File formats

• .fif — Elekta/Megin/MNE raw MEG.
• .vhdr / .eeg / .vmrk — BrainVision EEG (header / data / markers).
• .mat — MATLAB behavioral logs (stimuli, keystrokes, timing). Load with scipy.io.loadmat or MATLAB.

Loading the events (keystroke / word / sentence timings)

The Brain2Qwerty code release ships a studies package that reads these recordings and behavioural logs, aligns them, and emits a standardized event dataframe. Install the public libraries and let it download and build the events for you:

pip install neuralset neuralfetch

Clone the Brain2Qwerty repo and run the snippet from its brain2qwerty/ directory (or pip install -e . there first) so that import studies resolves — studies is the package that defines and registers Pinet2024Meg / Pinet2024Eeg; it is not part of neuralfetch.

import studies  # noqa: F401  - registers Pinet2024Meg / Pinet2024Eeg
from neuralset.events import Study

study = Study(name="Pinet2024Meg", path="SpanishBCBL")  # use "Pinet2024Eeg" for EEG
study.download()        # fetch this study's recordings + logs from this HF repo into `path`
events = study.build()  # standardized event dataframe across all subjects/sessions

events has one row per event, with type (Keystroke / Word / Sentence, plus the raw Meg/Eeg recording rows) and the timings in start (onset, seconds) and duration (seconds). If you have already downloaded the dataset, point path at the local folder and call study.build() directly (skip download()).

Ethics & privacy

Recordings are from consenting healthy adult volunteers under the study's approved ethics protocol at BCBL. Directly identifying material (structural MRI/T1, head-position videos, eye-tracking, and session videos) present in the internal dataset has been excluded from this public release; only de-identified M/EEG recordings and behavioral logs are included.

License

Released under CC BY-NC 4.0.

Citation

If you use this dataset, please cite:

@article{Lvy2026,
  title = {Noninvasive decoding of typed sentences from human brain activity},
  ISSN = {1546-1726},
  url = {http://dx.doi.org/10.1038/s41593-026-02303-2},
  DOI = {10.1038/s41593-026-02303-2},
  journal = {Nature Neuroscience},
  publisher = {Springer Science and Business Media LLC},
  author = {Lévy,  Jarod and Zhang,  Mingfang and Pinet,  Svetlana and Rapin,  Jérémy and Banville,  Hubert and d’Ascoli,  Stéphane and King,  Jean-Rémi},
  year = {2026},
  month = June 
}
@article{zhang2025thoughtactionhierarchyneural,
      title={From Thought to Action: How a Hierarchy of Neural Dynamics Supports Language Production}, 
      author={Mingfang Zhang and Jarod Lévy and Stéphane d'Ascoli and Jérémy Rapin and F. -Xavier Alario and Pierre Bourdillon and Svetlana Pinet and Jean-Rémi King},
      year={2025},
      eprint={2502.07429},
      archivePrefix={arXiv},
      primaryClass={q-bio.NC},
      url={https://arxiv.org/abs/2502.07429}, 
}

Acknowledgements

Supported by the Basque Government (BERC 2022–2025) and the Spanish State Research Agency (BCBL Severo Ochoa accreditation). Parts of this work were carried out within the European Union's Horizon 2020 programme under the Marie Skłodowska-Curie grant agreement No 945304 (Cofund AI4theSciences, PSL University).