Research map / Architectures and efficiency / RNNs and normalization
LibriBrain: Over 50 Hours of Within-Subject MEG to Improve Speech Decoding Methods at Scale
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 LibriBrain, a public 52-hour within-subject MEG dataset from one participant listening to Sherlock Holmes audiobooks, with a pnpl Python loader and baselines for speech detection, phoneme and word classification.
Problem
Non-invasive speech decoding lacks datasets with substantial depth per subject; most existing MEG/EEG speech datasets are broad but shallow (1-2 hours per subject), limiting within-subject scaling studies and standardized benchmarking.
Approach
A single-participant, deep MEG corpus (over 50 hours, 306-channel MEGIN system) with time-aligned word/phoneme annotations, held-out splits recorded on a separate day to prevent leakage, and a pip-installable loader plus baseline models for speech detection, phoneme and word classification.
Key results
- Speech detection: F1 0.8989 and balanced accuracy 0.7082 vs random 0.7848 and 0.4990 (10 seeds, p<0.01).
- Phoneme classification: micro F1 0.1168 vs random 0.0442; macro F1 showed no improvement over random (0.0253 vs 0.0258).
- Word classification top-10 balanced accuracy 0.3621 vs 0.3261 for the prior Armeni dataset (250-word vocabulary).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the MEG CNN model instead of a frequency-based random baseline | Speech detection · Speech detection | F1-score: 0.7848 ± 0.0017 -> 0.8989 ± 0.0007 | Helped |
| Use the MEG CNN model instead of a frequency-based random baseline | Speech detection · Speech detection | Balanced Accuracy: 0.4990 ± 0.0028 -> 0.7082 ± 0.0031 | Helped |
| Use the MEG CNN model instead of a frequency-based random baseline | Speech detection · Speech detection | Cross Entropy Loss: 0.5449 ± 0.0000 -> 0.3802 ± 0.0044 | Helped |
| Use the MEG CNN model instead of a frequency-based random baseline | Phoneme classification · Phoneme classification | Micro F1-score: 0.0442 ± 0.0003 -> 0.1168 ± 0.0003 | Helped |
| Use the MEG CNN model instead of a frequency-based random baseline | Phoneme classification · Phoneme classification | Macro F1-score: 0.0258 ± 0.0003 -> 0.0253 ± 0.0005 | Hurt |
| Use the MEG CNN model instead of a frequency-based random baseline | Phoneme classification · Phoneme classification | Balanced Accuracy: 0.0259 ± 0.0003 -> 0.0399 ± 0.0003 | Helped |
| Use the MEG CNN model instead of a frequency-based random baseline | Phoneme classification · Phoneme classification | Cross Entropy Loss: 3.3509 ± 0.0000 -> 3.2332 ± 0.0026 | Helped |
| Train on LibriBrain rather than the Armeni dataset | Word classification method · Word classification (top 250 words) | Top-10 Balanced Accuracy: 0.3261 ± 0.0019 -> 0.3621 ± 0.0031 | Helped |
| Increase the amount of training data | Speech detection and phoneme classification models · Speech detection and phoneme classification | Decoding performance | Helped |
| Average more instances of the same phoneme before classification | Phoneme classification · Phoneme classification | Balanced accuracy: exceeding 60% with 100 repetitions | Helped |
Surprising findings
- Macro F1 for phoneme classification did not beat the random baseline, attributed to a power-law phoneme frequency distribution.
- Averaging 100 phoneme repetitions was needed to exceed 60% balanced accuracy, suggesting single-trial phoneme decoding remains weak.
Limitations
- Single-subject, single-narrator, single-author design limits generalisability across people and language diversity.
- Listening paradigm rather than overt or covert speech; transfer to speech production is untested.
- Minimal preprocessing and no brain-to-text (sentence-level) baselines reported.
Experimental setup
Datasets: LibriBrain
Benchmarks: LibriBrain speech detection LibriBrain phoneme classification LibriBrain word classification
Metrics: F1-score balanced accuracy AUROC Jaccard index cross entropy loss top-10 balanced accuracy
Baselines: random baseline Armeni et al. word classification dataset/model
Builds on: SEANet SigLIP Kaldi phone position conventions
Closest papers by meaning
- LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale (NeurIPS 2026 · similarity 0.96)
- The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.87)
- Brain Treebank: Large-scale intracranial recordings from naturalistic language stimuli (NeurIPS 2024 · similarity 0.85)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.83)
- Du-IN: Discrete units-guided mask modeling for decoding speech from Intracranial Neural signals (NeurIPS 2024 · similarity 0.83)
- Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.