cs.LGAug 25, 2026
SaveLibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale
Organizations: PNPL, Department of Engineering Science, University of Oxford, UK
Abstract
We introduce LibriBrain100, a large-scale MEG dataset for speech decoding designed from the ground up for reproducible, standardised evaluation. LibriBrain100 more than doubles the size of the original LibriBrain release, resulting in over 100 hours of high-quality MEG acquired while subjects listened to naturalistic continuous speech. With 80 hours from a single subject, LibriBrain100 sets a new record for deep, within-subject neural data (8 more than the next comparable dataset and roughly 80 more than other datasets). To demonstrate the payoff of this depth-first design, we evaluate on a word-classification benchmark---an increasingly well-established stepping stone towards the open challenge of noninvasive brain-to-text decoding. Using an existing decoding model, we achieve state-of-the-art performance---validating both the quality of the recordings and the value of within-subject data at scale. Because collecting 80 hours of data per user is impractical for real-world applications, we also collected 40 minutes of additional data from each of 32 subjects. Using the same word-classification benchmark, we demonstrate the value of broad multi-subject data: supervised finetuning of a pre-trained model can substantially compensate for limited per-subject data. We provide standard train, validation, and test splits, all reproducible through an open-sourced Python library that supports easy downloading, optional preprocessing, and data loading for common deep learning frameworks. In addition, the dataset and evaluation infrastructure are being released alongside an open machine-learning competition with a public leaderboard for standardised benchmarking. Ultimately, our hope is that LibriBrain100 will accelerate progress towards practical non-invasive brain-computer interfaces, capable of restoring communication to people living with severe paralysis.
Figures & tables
Figure 1: Decoding words from LibriBrain100 . LibriBrain100 supports word-classification benchmarks across both deep within-subject data and broad cross-subject data. ( a ) Within-subject decoding : example speech stimuli and MEG responses from subject 0 across multiple corpora. Target words are highlighted in the text, aligned to the acoustic spectrogram, and marked in the corresponding MEG time series. MEG time series were averaged across temporal sensors illustrated in the sensor layout. ( b ) Cross-subject decoding : the same target-word classification task can be evaluated across the 32 additional subjects with limited per-subject data, enabling benchmarking for data-efficient generalisation.
| Component | Source | # Subj | All Data | Train | Val | Test |
| Deep | ||||||
| Subject 0 | Sherlock (all books) | 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 | |
| The Moth (30 podcasts) | 1 | 6.0 | 5.7 | 0.2 | 0.2 | |
| Subtotal (hours) | 1 | 80.5 | 78.6 | 1.0 | 0.9 |
Table 1: LibriBrain100 data sources and standard splits . The dataset comprises a deep single-subject component ( 80 hours) and a broad multi-subject component ( 40 minutes per subject across 32 subjects). All durations are shown in hours and rounded to one decimal place; totals are computed before rounding. For a visual summary of the dataset, see Figure 5 (Appendix A.1 ).
Figure 2: Subject 0 performance on held-out test data . Performance improves when training jointly with Subject 1-32 training data, except on the Sherlock subset. Here, training only on Sherlock training data leads to the best performance. Random chance accuracy is . * indicates and ** indicates under Mann-Whitney U-tests.
Figure 3: Fine-tuning MEG-XL on subjects 1-32, with and without training alongside Subject 0 . We report performance on the Sherlock test set. Training jointly with Subject 0 significantly improves generalisation on the broad and shallow data of subjects 1-32. Random chance accuracy is . *** indicates under a Mann-Whitney U-test. The per-subject results in the right panel illustrate that difference is not driven by any specific subject, or proper subset of subjects, but is rather consistent across all 32 subjects.
Figure 4: Performance as a function of fine-tuning data size . Generalisation remains robust, even with only 25% training data, equivalent to about 10 minutes of brain recordings (i.e., a small enough amount of data to be interesting for clinical applications). The largest difference comes from including data from Subject 0. Random chance accuracy is . None of the differences in performance under different training data percentages were significant under Mann-Whitney U-tests.
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Dataset proportions .( a ) Recording hours by subject and corpus. The stacked bar plot shows total duration per subject, with sub-0 comprising multiple linguistic materials and other subjects a subset of audiobook data. The bubble plot represents the same information (area proportional to hours); the larger sub-0 circle is subdivided into audiobooks, sentences, and podcasts, while smaller circles represent other subjects. Colours indicate data type, blue shades distinguish between audiobooks, and circle outlines distinguish between subjects. Black concentric circles provide a size reference (~1 h, ~10 h, ~80 h). ( b ) Treemap of the full dataset, with rectangle area proportional to duration. Categories are subdivided into corpora and then into MEG recording sessions. The multicolour outline marks the only two audiobook sessions recorded from multiple subjects.
| Name | Dataset | Language | Sensors | Total hrs | # Subj | Hrs/Subj | Stimulus | Public | PNPL Dataloader |
| MOUS | Schoffelen et al. (2019) | Dutch | 275 | 81 | 96 | 0.8 | Spoken sentences | Yes | Yes |
| MEG-MASC | Gwilliams et al. (2023) | English | 208 | 49 | 27 | 1.0 | MASC stories (synthetic voice) | Yes | Yes |
| Armeni | Armeni et al. (2022) | English | 269 | 30 | 3 | 10.0 | LibriVox (Sherlock Book 3) | Yes | Yes |
| Le Petit Prince | d’Ascoli et al. (2025) | French | 306 | 93 | 58 | 1.6 | Le Petit Prince | Yes | Yes |
| LibriBrain | Özdogan et al. (2025) | English | 306 | 52 | 1 | 52.3 | LibriVox (Sherlock Book 1–7) | Yes | Yes |
| LibriBrain100 | (ours) | English | 306 | 104 | 33 | 0.6– 80.0 | LibriVox (Sherlock Book 1–9), TIMIT, MOCHA-TIMIT, The Moth (30 Podcasts) | Yes | Yes |
Table 2: Breadth and depth of speech comprehension MEG datasets .
| Book | Name | Type | Sessions | Hours |
| 1 | A Study in Scarlet (1888) | Novel | 14 | 04:37:34 |
| 2 | The Sign of the Four (1890) | Novel | 12 | 04:27:31 |
| 3 | The Adventures of Sherlock Holmes (1892) | Short Stories | 12 | 10:56:13 |
| 4 | The Memoirs of Sherlock Holmes (1893) | Short Stories | 12 | 08:53:17 |
| 5 | The Hound of the Baskervilles (1901–1902) | Novel | 15 | 06:10:32 |
| 6 | Return of Sherlock Holmes (1905) | Short Stories | 14 | 11:51:17 |
Table 3: Audiobooks available in LibriBrain100 . Books 1–7 appeared in the original LibriBrain release ( Özdogan et al., 2025 ) ; the rest are new here. Text in the name column link to the source audio on LibriVox.
| # | Story | Author | Duration |
|---|---|---|---|
| 1 | Alternate Ithaca Tom | Tom Weiser | 11:47 |
| 2 | Thumbs Up! | Nathan Englander | 14:09 |
| 3 | Goldie, the Goldfish | Becca Stevens | 10:54 |
| 4 | Under the Influence | Jeffery Rudell | 10:28 |
| 5 | Treasure Island | Boots Riley | 13:30 |
| 6 | That Thing on My Arm | Padma Lakshmi | 14:49 |
Table 4: Podcasts available in LibriBrain100 . Story titles are hyperlinked to the source audio and transcript at https://themoth.org/stories/ . Validation and test set stories are highlighted yellow and red , respectively. Durations are given in minutes and seconds for the audio files used in our stimulus set.
| Test set (24 speakers) | ||||
| mdab0 | mwbt0 | felc0 | mtas1 | mwew0 |
| fpas0 | mjmp0 | mlnt0 | fpkt0 | mlll0 |
| mtls0 | fjlm0 | mbpm0 | mklt0 | fnlp0 |
| mcmj0 | mjdh0 | fmgd0 | mgrt0 | mnjm0 |
| fdhc0 | mjln0 | mpam0 | fmld0 | |
| Validation set (50 speakers, from dev split) | ||||
Table 5: TIMIT speaker IDs used for the test and validation splits .
Figure 6: Speaker embeddings . Two-dimensional t-SNE projection of ECAPA-TDNN speaker embeddings, which capture acoustic properties such as timbre, pitch, and vocal tract characteristics. Each point represents a single speaker, obtained by averaging embeddings across multiple utterances. Marker fill colour indicates the speech corpus (TIMIT, Podcasts, MOCHA-TIMIT, Audiobook), while marker edge colour denotes speaker sex (male, female).
Figure 7: Phoneme frequency distribution per corpus . Bar plots show normalised phoneme frequencies for each corpus, sorted by decreasing frequency in TIMIT. Each phoneme category on the x-axis contains three adjacent bars corresponding to TIMIT (orange), Podcasts (green), and Audiobook (blue). The y-axis is displayed on a logarithmic scale to enhance visibility of low-frequency phonemes. Selected low-frequency phonemes (G, Y, SH, OY, ZH) are highlighted with downward-pointing arrows and thicker bar outlines to facilitate comparison across corpora.
Figure 8: Phoneme rank-frequency distribution per corpus . Each point represents a phoneme, with its position determined by its rank (x-axis) and normalised frequency (y-axis), both shown in logarithmic scale. Phonemes are ranked in decreasing order of frequency based on TIMIT. Dashed lines indicate linear fits computed over ranks. Different marker shapes (circles, squares, triangles) and different colors (orange, green, blue) distinguish the three corpora.
Figure 9: Keyword embeddings per corpus . Two-dimensional t-SNE projection of keyword embeddings from the Audiobook (blue), TIMIT (orange), and Podcasts (green) corpus . Each point corresponds to a keyword, and proximity between points reflects similarity in the embedding space, capturing lexical-semantic relationships between words.
Figure 10: MEG evoked response within subject . Differences between speech and non-speech segments are shown as a proxy for auditory and linguistic processing. ( a ) Topographic map of the root mean square (RMS) difference between speech and non-speech evoked responses, averaged across sessions, showing stronger amplitude for speech over bilateral temporal sensors. Statistically significant sensors (p<0.001) are marked with white dots. ( b ) Time–frequency representation of the difference in inter-trial coherence (ITC) between speech and non-speech segments, averaged across sessions. Speech elicits greater phase consistency in low-frequency bands (~2–7 Hz; delta–theta range). Statistically significant clusters (p<0.001) are outlined with a black contour. ( c ) Power spectral density (PSD) for speech (red) and non-speech (blue) segments, averaged across sessions, indicating increased low-frequency power during speech. Statistically significant frequency ranges (p<0.001) are highlighted with a semi-transparent grey region. For each panel, data are shown for a single representative participant (sub-0), averaged across recording sessions, where each session corresponds to one chapter (1-14) of A Study in Scarlet from Sherlock Holmes audiobook.
Figure 11: MEG evoked response between subjects . Data are shown for all participants (sub-0 to sub-32), averaged across subjects, for chapters 11–12 of Study in Scarlet from Sherlock Holmes audiobook. Differences between speech and non-speech segments are shown as a proxy for auditory and linguistic processing. ( a ) Topographic map of the root mean square (RMS) difference between speech and non-speech evoked responses, averaged across subjects, showing stronger amplitude for speech over bilateral temporal sensors. Statistically significant sensors (p<0.001) are marked with white dots. ( b ) Time–frequency representation of the difference in inter-trial coherence (ITC) between speech and non-speech segments, averaged across subjects. Speech elicits greater phase consistency in low-frequency bands ( 2–7 Hz; delta–theta range). Statistically significant clusters (p<0.001) are outlined with a black contour. ( c ) Power spectral density (PSD) for speech (red) and non-speech (blue) segments, averaged across subjects, indicating increased low-frequency power during speech. Statistically significant frequency ranges (p<0.001) are highlighted with a semi-transparent grey region.
Figure 12: Variability across recording sessions . Spatial, temporal, and spectral components showing strong discrimination between speech and non-speech segments are illustrated across recording sessions to assess variability in amplitude, phase, and power. ( a ) Amplitude is quantified as the root mean square (RMS) of the evoked response. ( b ) Phase consistency is measured using inter-trial coherence (ITC). ( c ) Oscillatory power is measured using power spectral density (PSD). For each metric, the mean is computed within functionally relevant masks derived from statistical testing on aggregate data, shown as grey-shaded insets in each panel: a sensor mask for amplitude, a time–frequency mask for phase, and a frequency mask for power. Error bars indicate variability within the corresponding mask (across sensors, time–frequency points, or frequencies, depending on the panel). Data are shown for a single representative participant (sub-0), with each bar corresponding to one recording session, i.e., one chapter (1–14) of A Study in Scarlet from Sherlock Holmes audiobook.
Figure 13: Variability across subjects . Spatial, temporal, and spectral components showing strong discrimination between speech and non-speech segments are illustrated across participants to assess variability in amplitude, phase, and power. ( a ) Amplitude is quantified as the root mean square (RMS) of the evoked response. ( b ) Phase consistency is measured using inter-trial coherence (ITC). ( c ) Oscillatory power is measured using power spectral density (PSD). For each metric, the mean is computed within functionally relevant masks derived from statistical testing on aggregate data, shown as grey-shaded insets in each panel: a sensor mask for amplitude, a time–frequency mask for phase, and a frequency mask for power. Error bars indicate variability within the corresponding mask (across sensors, time–frequency points, or frequencies, depending on the panel). Data are shown across participants (sub-0 to sub-32), with each bar corresponding to one subject. Colours differentiate participants.
Figure 14: d’Ascoli (Supervised) vs MEG-XL (Self-supervised). We compare the supervised word decoding model proposed by d’Ascoli et al. (2025) to fine-tuning the pre-trained self-supervised MEG-XL model proposed by Jayalath and Parker Jones (2026) . On deep Subject 0 data, the supervised model performs better, while on the shallow data of subjects 1-32, MEG-XL generalises better owing to its data-efficiency. Random chance is . ** indicates under a Mann-Whitney U-test.
Figure 15: Information transfer comparison. We measure the OVMI achieved by Moses et al. (2021) , calculating the quantity from their reported decoder accuracy and 50-word vocabulary, against MEG-XL fine-tuned on LibriBrain with our target vocabulary and against an OVMI-optimised vocabulary. We use SUBTLEX-UK as the communication distribution that parameterises OVMI. Higher OVMI scores are better. The results here are provided as references for future benchmarks.
| # | Word | # | Word | # | Word | # | Word | # | Word |
| 1 | is | 11 | he | 21 | but | 31 | at | 41 | do |
| 2 | the | 12 | that | 22 | will | 32 | out | 42 | can |
| 3 | a | 13 | have | 23 | so | 33 | our | 43 | time |
| 4 | to | 14 | this | 24 | all | 34 | am | 44 | think |
| 5 | it | 15 | they | 25 | my | 35 | it’s | 45 | good |
| 6 | i | 16 | of | 26 | for | 36 | had | 46 | always |
Table 6: The 50-word vocabulary used in the word classification task .
Figure 16: Model predictions on full vocabulary . ( a ) Top-10 accuracy across all word occurrences in Chapters 11 and 12 of A Study in Scarlet and across all 32 subjects, grouped by Part-of-Speech (POS) category (purple function, red adverb, green adjective, orange verb, blue noun). Predictions were evaluated against the full vocabulary, and accuracy is defined as the proportion of events for which the true word appears within the top-10 predicted words. ( b ) Word cloud of the 100 best-predicted words, ranked by top-10 balanced accuracy. Word size is proportional to top-10 balanced accuracy, and word colours indicate their corresponding Part-of-Speech categories.