Skip to content
Open Imaging Index
MRI Brain

Cephalonauts One

Cephalonauts One: a deep fMRI dataset for decoding naturalistic speech in the human brain

Brain fMRI of healthy adults listening to French podcasts for about 30 hours each, recorded to train and benchmark models that decode heard speech from brain activity. 3 native French speakers, 72 sessions and 358 runs at 3 T, with time-aligned transcripts and audio and text features.

Go to dataset Edit Report

Overview

Cephalonauts One is a deeply sampled fMRI dataset for brain decoding of natural speech. The company Karavela recorded three healthy volunteers while they listened to French radio podcasts over many sessions, so that models can learn to recover what a person heard from their brain activity alone. The release comes with fixed training, validation and test sessions and baseline code, and it was accepted at the NeurIPS 2026 Evaluations and Datasets track.

Composition

Three native French speakers aged 31 to 51 took part, with about 30 hours of functional MRI each: 24 sessions per person, 358 runs and 89.31 hours in total. The stimuli are unscripted stories from a French public radio podcast. Each run comes with the audio, a French transcript aligned word by word to the scans, and audio and text embeddings from pretrained models aligned to the same time points. Per subject, held-out sessions serve as validation and test data.

Acquisition

All sessions were recorded on one 3 T Siemens MAGNETOM Cima.X scanner with a 64-channel head and neck coil at the Paris Brain Institute. The functional scans use a fast multiband BOLD sequence with a repetition time of 520 ms at 3 mm resolution, which gives about 207,000 volumes per subject. A T1-weighted anatomical scan was also acquired, and the data were preprocessed with DeepPrep. The release offers the functional data in native anatomical space, in a standard template space and on the fsaverage5 cortical surface.

Annotations

There are no image labels. The targets for decoding are the time-aligned transcripts and the audio and text features.

Known limitations

With three participants, the data suit models trained per person, not conclusions about a population. Sex and other demographics beyond the age range are not reported. The public file list holds brain masks but no raw anatomical images. The terms of use limit use to non-commercial scientific research and forbid moving the data outside the European Economic Area without adequate protection under the GDPR.

Cohort

Aggregate numbers from the sources below. Bars are relative to the 3 subjects.

Age

range 31 to 51

No age bins reported.

Section 3.1 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

Condition

Groups can overlap

  • Healthy control 3 100%

Section 3.1 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

Field strength

  • 3 T 3 100%

Section 3.2 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

Contrast / sequence

Groups can overlap

  • BOLD fMRI 3 100%

Table 2 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

Scanner vendor

  • Siemens Healthineers 3 100%

Section 3.2 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

Country

  • France 3 100%

Section 3.2 in Collas et al. 2026, Cephalonauts One, a deep fMRI dataset for decoding naturalistic speech (arXiv:2610.03558v1)

License and access

Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer

Access
Signed agreement

Sign a data use agreement, often reviewed by the provider

Log in to Hugging Face, fill in a form with affiliation and intended use, and accept the Cephalonauts One terms of use. Karavela approves requests manually.

Access page

Cephalonauts One terms of use

Non-commercial scientific research only, after accepting the terms on Hugging Face and manual approval. No re-identification, no transfer outside the EEA without adequate GDPR protection, data deleted after use. Data and derived data may be passed on only under the same terms, and publications must carry a fixed acknowledgement.

The arXiv record of the paper is CC BY 4.0. That covers the paper only; the data are under the gated terms of use.

Original license text Version read: Gated access terms on the Hugging Face dataset card, revision of 2026-10-13 Checked 2026-10-23

What you can do

  • No
  • Not stated
  • Conditional
  • Conditional
  • Conditional

What you can share

  • Conditional
  • Conditional
  • Not stated

What you must do

  • Yes
  • Yes
  • Yes
  • Conditional
  • Manuscript review No
  • Release code No
  • No
  • Yes

Limits

  • Yes
  • Conditional

Commercial license: Not stated The terms mention no commercial license.

Citation

Collas A, Jalouzot L, Ilinca G, et al. Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain. Advances in Neural Information Processing Systems (NeurIPS) 2026, Evaluations and Datasets Track. arXiv:2610.03558

@inproceedings{collas2026cephalonauts,
  title = {{Cephalonauts One}: A deep {fMRI} dataset for decoding naturalistic speech in the human brain},
  author = {Collas, Antoine and Jalouzot, Louis and Ilinca, Géraud and Caris, Corentin and Valabrègue, Romain and Hassayoune, Ahmed and Goncalves, David and Hueber, Madeleine and Delebarre, Thaddée and Savatovsky, Julien and Fonteneau, Clara and Maussion, Charles and Thirion, Bertrand and Thual, Alexis},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026},
  note = {Evaluations and Datasets Track},
  url = {https://arxiv.org/abs/2610.03558}
}

Sources

Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.