A new open dataset of brain recordings links two kinds of training data: many hours from one person and shorter recordings from a wider group. In a fixed 50-word word-classification benchmark, including the deep-recording subject in training was associated with higher performance when the model was tested on data from 32 other people.
The clearest difference appeared on a Sherlock test set. Balanced top-10 accuracy gives each word class equal weight and counts a prediction when the correct word is among the model’s ten guesses. The score was 0.478 when training included Subject 0 and 0.329 when it did not, a gap of about 15 percentage points. The paper marked the difference at p < .001 under a Mann-Whitney U-test.
A dataset built around two scales
The release contains 104.2 hours of magnetoencephalography, or MEG, data. Of that total, 80.5 hours come from a deep component recorded from Subject 0, while 23.7 hours come from a broad component involving subjects 1 to 32. Across the standard partitions, the dataset contains 78.6 training hours, 12.5 validation hours and 13.1 test hours.
The data use standard training, validation and testing partitions. For the Sherlock material, session 11 is held out for validation and session 12 for testing.
The recordings came from 33 healthy volunteers, including 10 women and 23 men aged 19 to 50, with a median age of 28. Sixteen participants were native English speakers and 17 were highly proficient second-language English speakers. All reported normal hearing, and none had a history of neurological disorders.
Participants passively listened to continuous speech drawn from Sherlock Holmes audiobooks, the TIMIT and MOCHA-TIMIT speech corpora, and The Moth podcasts. The recordings used 102 magnetometers and 204 planar gradiometers. After filtering and downsampling, the released data had 306 channels sampled every 4 milliseconds.
The gains were not uniform
The researchers compared performance on Subject 0 with and without training data from subjects 1 to 32. Including the broader-subject data was generally associated with higher test performance for TIMIT, MOCHA-TIMIT and podcast material. The Sherlock subset went slightly the other way, performing marginally better when the additional subjects were left out.
The difference between the joint-training and no-Subject-0 conditions remained when the amount of supervised fine-tuning data was reduced. With Subject 0 included, balanced top-10 accuracy was 0.492 using all available data, 0.488 using half and 0.482 using a quarter. Without Subject 0, the corresponding scores were 0.337, 0.334 and 0.332. Differences across the data fractions were not statistically significant, and the quarter-data condition represented about 10 minutes of material.
Model choice changed with the data
A supplementary comparison showed that model choice mattered. On the deep Subject 0 data, a supervised model scored 0.739, compared with 0.585 for the pre-trained MEG-XL model. On the broad data from subjects 1 to 32, the pattern reversed: MEG-XL scored 0.478, compared with 0.328 for the supervised model. The figure marked these differences at p < .01.
An auxiliary information-transfer comparison reported 0.259 bits per word attempt for the Moses comparator, 0.075 for MEG-XL with the 50-word target vocabulary and 0.147 with an optimised 32-word vocabulary. The authors describe these as reference values and caution that the recordings represent heard speech rather than speech participants were trying to produce.
In a between-subject proxy analysis, speech differed from non-speech in several group-level neural measures. Speech was associated with stronger bilateral temporal RMS, greater low-frequency inter-trial coherence around 2 to 7 hertz and increased low-frequency power. Figure captions marked these features at p < .001, although the paper also reports substantial variation between participants.
A benchmark, not a finished communicator
The evaluation concerns word classification with a fixed 50-word vocabulary, based on passive listening. It does not show decoding of attempted or imagined speech, open-ended brain-to-text communication or performance in clinical populations. The authors present the release as shared infrastructure for neural speech-decoding research while acknowledging the gap between passive-listening benchmarks and real communication.
The release includes raw BIDS recordings in FIF format and minimally preprocessed float32 HDF5 data. The accompanying pnpl Python library is described as supporting task-driven loading, time-locked annotations and on-demand downloads from Hugging Face. The paper identifies the work as arXiv preprint 2608.25204v1, dated 25 August 2026.
The training-condition comparisons show associations between data composition and benchmark performance, not proof that one person’s recordings cause better decoding. Further testing will be needed to determine how well the reported transfer generalises to new speakers, stimuli, vocabularies, attempted or imagined speech and clinical populations.
Paper data and sources
Original title: LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale
Authors: Francesco Mantegna, Dulhan Jayalath, Gereon Elvers et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text