wordcab/libribrain-meg-preprocessed

LibriBrain MEG Preprocessed Dataset

Preprocessed magnetoencephalography (MEG) recordings with phoneme labels from the LibriBrain dataset, optimized for fast loading during machine learning model training.

This dataset was created for the LibriBrain 2025 Competition (now concluded).

Dataset Overview

MEG Recording Specifications

• Channels: 306 total (102 magnetometers + 204 gradiometers)
• Sampling Rate: 250 Hz
• Duration: ~52 hours of recordings
• Subject: Single English speaker listening to Sherlock Holmes audiobooks
• Phoneme Instances: ~1.5 million

Phoneme Inventory

39 ARPAbet phonemes with position encoding:

• Vowels (15): aa, ae, ah, ao, aw, ay, eh, er, ey, ih, iy, ow, oy, uh, uw
• Consonants (24): b, ch, d, dh, f, g, hh, jh, k, l, m, n, ng, p, r, s, sh, t, th, v, w, y, z, zh
• Special: oov (out-of-vocabulary)

Position markers: B (beginning), I (inside), E (end), S (singleton)

Signal Processing

All MEG data has been preprocessed through the following pipeline:

1. Bad channel removal
2. Signal Space Separation (SSS) for noise reduction
3. Notch filtering for powerline noise removal
4. Bandpass filtering (0.1-125 Hz)
5. Downsampling to 250 Hz

Preprocessing and Grouping

This dataset contains pre-grouped and averaged MEG samples for significantly faster data loading during training. Instead of grouping samples on-the-fly (which is computationally expensive), samples have been pre-grouped at various levels.

Available Grouping Configurations

• grouped_5: 5 samples averaged together
• grouped_10: 10 samples averaged together
• grouped_15: 15 samples averaged together
• grouped_20: 20 samples averaged together
• grouped_25: 25 samples averaged together
• grouped_30: 30 samples averaged together
• grouped_35: 35 samples averaged together (partial - train only)
• grouped_45: 45 samples averaged together
• grouped_50: 50 samples averaged together
• grouped_55: 55 samples averaged together
• grouped_60: 60 samples averaged together
• grouped_100: 100 samples averaged together

Each configuration contains:

• train_grouped.h5: Training data
• validation_grouped.h5: Validation data
• test_grouped.h5: Test data
• paths.yaml: File path references

Why Use Grouped Data?

• Faster Loading: Pre-computed grouping eliminates runtime averaging overhead
• Memory Efficient: Smaller file sizes for higher grouping levels
• Flexible: Choose grouping level based on your accuracy vs. speed requirements
• Standardized: Consistent preprocessing across all configurations

Installation

This dataset requires the modified pnpl library for loading:

pip install git+https://github.com/September-Labs/pnpl.git

Usage

from pnpl.datasets import GroupedDataset

# Load preprocessed data with 100-sample grouping
train_dataset = GroupedDataset(
    preprocessed_path="data/grouped_100/train_grouped.h5",
    load_to_memory=True  # Optional: load entire dataset to memory for faster access
)

val_dataset = GroupedDataset(
    preprocessed_path="data/grouped_100/validation_grouped.h5",
    load_to_memory=True
)

# Get a sample
sample = train_dataset[0]
meg_data = sample['meg']        # Shape: (306, time_points)
phoneme_label = sample['phoneme']  # Phoneme class index

# Use with PyTorch DataLoader
from torch.utils.data import DataLoader

dataloader = DataLoader(
    train_dataset,
    batch_size=32,
    shuffle=True,
    num_workers=4
)

Data Structure

data/
├── grouped_5/
│   ├── train_grouped.h5
│   ├── validation_grouped.h5
│   ├── test_grouped.h5
│   └── paths.yaml
├── grouped_10/
│   ├── train_grouped.h5
│   ├── validation_grouped.h5
│   ├── test_grouped.h5
│   └── paths.yaml
├── ...
└── grouped_100/
    ├── train_grouped.h5
    ├── validation_grouped.h5
    ├── test_grouped.h5
    └── paths.yaml

File Sizes

Grouping	Train	Validation	Test	Total
grouped_5	45.6 GB	425 MB	456 MB	~47 GB
grouped_10	22.8 GB	213 MB	228 MB	~24 GB
grouped_20	11.4 GB	106 MB	114 MB	~12 GB
grouped_50	4.6 GB	37 MB	42 MB	~4.7 GB
grouped_100	2.3 GB	19 MB	21 MB	~2.4 GB

Dataset Splits

• Train: 88 sessions (~51 hours)
• Validation: 1 session (~0.36 hours)
• Test: 1 session (~0.38 hours)

Citation

If you use this dataset, please cite the LibriBrain competition:

@misc{libribrain2025,
  title={LibriBrain: A Dataset for Speech Decoding from Brain Signals},
  author={Neural Processing Lab},
  year={2025},
  url={https://neural-processing-lab.github.io/2025-libribrain-competition/}
}

License

Please refer to the original LibriBrain dataset license terms.

Acknowledgments

This preprocessed version was created to facilitate faster training for the LibriBrain 2025 Competition. The original dataset and competition were organized by the Neural Processing Lab.