Skip to content
The BCI Briefing

Dataset3 min read

LibriBrain100 turns cross-subject speech decoding into an open competition

The largest open MEG speech dataset now adds 32 subjects, shifting the benchmark from how well a decoder learns one brain to how quickly it adapts to a new one.

ByThe BCI Briefing

Evidence class: Reported claim

The short version

  • LibriBrain100 adds 32 subjects with about 40 minutes each, plus about 80 hours from the original subject.
  • The 2026 competition targets word classification and efficient cross-subject generalisation.
  • The data are non-invasive MEG recorded while listening to speech, not attempted speech.

The organisers of the PNPL competition described its 2026 edition in a preprint posted on 3 September. The 2025 edition, built on roughly 50 hours of MEG from a single subject, produced winning F1-macro scores of 95.6% for speech detection and 73.6% for phoneme classification.

The new release, LibriBrain100, extends the within-subject recordings to about 80 hours and adds 32 further subjects with around 40 minutes each. The tasks move up in linguistic complexity to word classification, and add a track on generalising to new subjects from minutes of data.

How solid is this

Preprint on arXiv; not yet peer reviewed. The reported 2025 scores are for detecting and classifying heard speech, which is easier than decoding intended speech. MEG requires a magnetically shielded room, so results do not transfer directly to wearable devices.

Preprint. A full manuscript shared publicly before peer review. Methods are open to scrutiny, but conclusions may change.

What to watch

  1. 01Competition results on the cross-subject track.
  2. 02Tests of whether models trained on speech perception help decode attempted or imagined speech.

Sources

  1. 01
  2. 02

    github.com/mne-tools/mne-python (opens in a new tab)

    Open-source repositoryGitHubChecked against source 3 Oct 2026

    Repository metadata read from the GitHub API.

  3. 03

    github.com/braindecode/braindecode (opens in a new tab)

    Open-source repositoryGitHubChecked against source 3 Oct 2026

    Repository metadata read from the GitHub API.