LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale
Francesco Mantegna, Dulhan Jayalath, Gereon Elvers, Tasha Kim, Benjamin Ballyk, Alex Fung, SungJun Cho, Teyun Kwon, Luisa Kurth, Miran Özdogan, Gilad Landau, Pratik Somaiya, Natalie Voets, Mark Woolrich, Oiwi Parker Jones
cs.LG, cs.CL
2026-08-26
LibriBrain100: 104 h heard-speech MEG, 80.5 h from one subject plus ~44 min from 32 others. The deep subject lifts 50-word top-10 from 0.329 to 0.478; ~10 min fine-tuning holds.
Invasive speech BCIs have pushed word error rate below 5% in some paralysed patients. They require surgery. Non-invasive methods are the path to scale, and they have lacked ImageNet-style infrastructure: shared data, standard splits, a public board. Oxford PNPL's earlier LibriBrain got one listener to 52.3 hours of MEG and decoding moved; it was still one subject and almost entirely Sherlock Holmes audiobooks. Depth without breadth.
LibriBrain100 adds both. Heard speech is not the BCI endgame (users need to attempt or imagine speech), but alignment is clean and SNR is relatively high, so it is a usable step.
Recordings were taken at Oxford OHBA on a MEGIN TRIUX Neo, 306 channels (102 magnetometers, 204 gradiometers), 1 kHz native, downsampled to 250 Hz. Thirty-three healthy volunteers, 10 female and 23 male, ages 19–50, median 28. Sixteen native English speakers; seventeen highly proficient L2.
Hours from Table 1:
| Split | Stimulus | Subjects | Total | Train / val / test |
| Deep, subject 0 | Sherlock canon | 1 | 68.1 | 67.3 / 0.4 / 0.4 |
| TIMIT | 1 | 5.3 | 4.9 / 0.2 / 0.2 | |
| MOCHA-TIMIT | 1 | 1.1 | 0.8 / 0.2 / 0.2 | |
| 30 Moth podcasts | 1 | 6.0 | 5.7 / 0.2 / 0.2 | |
| Deep subtotal | 1 | 80.5 | 78.6 / 1.0 / 0.9 | |
| Broad, subjects 1–32 | two Sherlock chapters | 32 | 23.7 | n/a / 11.5 / 12.2 |
| Total | 33 | 104.2 | 78.6 / 12.5 / 13.1 |
Within-subject depth is about 8× Armeni (3 × 10 h) and about 80× the other usual MEG listening sets. TIMIT and MOCHA-TIMIT control phoneme statistics for comparison with ASR and invasive BCI work. Thirty Moth stories add semantic range, matching Tang et al.'s fMRI semantic-decoding stimuli.
The headline task is 50-word classification: a MEG window after word onset mapped to a closed vocabulary, scored as class-averaged top-10 accuracy. Chance is 0.20. The backbone is MEG-XL, self-supervised on about 300 hours from 800 subjects, then fine-tuned here. Data live on Hugging Face; the loader is pnpl.
Held-out subject 0 (Figure 2, balanced top-10):
| Test subset | S0+S1–32 | S0 only | S0 Sherlock only |
| Sherlock audiobook | 0.492 | 0.519 | 0.585 |
| TIMIT | 0.431 | 0.393 | n/a |
| MOCHA-TIMIT | 0.501 | 0.436 | n/a |
| Podcasts | 0.464 | 0.399 | n/a |
In-domain Sherlock prefers the narrowest training mix. On TIMIT, MOCHA, and podcasts, the 32 shallow subjects help subject 0 generalise.
Cross-subject Sherlock (Figure 3): adding subject 0's 80 hours lifts the 32-person aggregate from 0.329 to 0.478, about 15 points, in the same direction for every subject. Chance remains 0.20.
Cut fine-tuning to 25% (10 minutes, Figure 4): with subject 0, 100% / 50% / 25% land at 0.492 / 0.488 / 0.482, differences not significant. Without subject 0 the numbers sit near 0.33. Transfer from the deep subject is doing most of the work; minutes of per-person fine-tuning are not the bottleneck.
In the appendix, a purely supervised d'Ascoli model reaches 0.739 on the deep subject, above MEG-XL's 0.585. On the shallow 32, MEG-XL wins 0.478 to 0.328. That comparison is not in the main figures.
Non-invasive speech decoding now has shared plumbing: splits, a Python loader, a planned public competition. Depth-first still pays: 80 hours from one listener can carry people who only have 40 minutes, or 10.
Against current MEG sets this is the deepest listening corpus, third in subject count (33; MOUS 96, Le Petit Prince 58). ML groups can skip FIF/BIDS and train on HDF5 windows. People who want their own preprocessing still get raw.
This is a dataset-plus-baseline paper, not a new decoder. Fifty words, heard speech, top-10: a long way from open brain-to-text.
The authors flag two remaining holes. All data are passive listening; a BCI needs attempted or inner speech, where SNR and alignment get worse. Inner-speech collection is underway, not in this release. They also withheld a brain-to-text baseline, judging methods too unstable to anchor a community number.
Subjects are healthy volunteers, so the 10-minute fine-tune story has no patient data behind it. The 50-word list is mostly function words; top-10 chance is already 0.20. MEG needs a shielded room and a multi-million-dollar scanner, not a headset EEG.
On the deepest subject, MEG-XL loses to the simpler supervised model in the appendix. The abstract's "state-of-the-art with an existing model" is doing a lot of work with an incomplete comparison.