Research map / Label-efficient and robust learning / Recognition and long-tail

LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Multilingual language representation pretraining · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces LibriBrain100, a 104-hour MEG speech dataset (about 80 hours from one subject plus 40 minutes from each of 32 others) with standard splits, a pnpl Python library, and word-classification baselines using fine-tuned MEG-XL.

Problem

Non-invasive neural speech decoding lacks large, standardised, reproducible MEG datasets that cover both deep within-subject data and cross-subject generalisation.

Approach

Extends the LibriBrain MEG dataset in depth (about 80 hours from one subject, adding TIMIT, MOCHA-TIMIT and Moth podcasts) and breadth (32 new subjects with about 40 minutes each), with standard splits, a Python loader, and word-classification benchmarks run with the pre-trained MEG-XL model fine-tuned on the data.

Key results

• Fine-tuned MEG-XL reaches state-of-the-art word classification on subject 0 test data when trained on the full ~80-hour within-subject data (per the abstract).
• Including subject 0 training data improves cross-subject generalisation for subjects 1-32 by about 15 percentage points.
• With only 25% of the ~40-minute per-subject fine-tuning data (about 10 minutes), generalisation remains robust with minimal degradation.

Surprising findings

• For subject 0 on the Sherlock test subset, training without subjects 1-32 gave the best (though marginal) performance.
• Reducing per-subject fine-tuning data to about 10 minutes causes minimal degradation, suggesting cross-subject transfer drives most performance.

Limitations

• Passive listening paradigm only; no speech production or imagined speech data yet.
• No brain-to-text baseline included in this release.
• All participants are healthy volunteers, so clinical generalisation is untested.

Experimental setup

Datasets: LibriBrain100 LibriBrain MOUS Cam-CAN SMN4Lang TIMIT MOCHA-TIMIT The Moth podcasts

Benchmarks: LibriBrain100 word classification (50-word vocabulary)

Metrics: top-10 balanced accuracy

Baselines: MEG-XL

Builds on: LibriBrain MEG-XL

Closest papers by meaning

• LibriBrain: Over 50 Hours of Within-Subject MEG to Improve Speech Decoding Methods at Scale (NeurIPS 2025 · similarity 0.96)
• The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.91)
• Brain Treebank: Large-scale intracranial recordings from naturalistic language stimuli (NeurIPS 2024 · similarity 0.88)
• Time-Masked Transformers with Lightweight Test-Time Adaptation for Neural Speech Decoding (NeurIPS 2025 · similarity 0.86)
• Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · similarity 0.85)
• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.85)