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

Neural speech decoding has reached an inflection point where progress is limited by the lack of standardized, large-scale datasets. While deep learning has transformed fields like computer vision through massive benchmarks, non-invasive brain-to-text decoding lacks similar infrastructure. This paper introduces LibriBrain100, a dataset designed to address the trade-off between data depth (hours per subject) and breadth (number of subjects) to accelerate the development of non-invasive brain-computer interfaces (BCIs).

LibriBrain100 comprises over 100 hours of magnetoencephalography (MEG) recordings from 33 subjects. The dataset follows a dual-design strategy:

• Depth: Approximately 80 hours of data from a single subject (Subject 0) listening to diverse stimuli, including the full Sherlock Holmes canon, phonetically controlled corpora (TIMIT and MOCHA-TIMIT), and semantically rich podcast narratives.
• Breadth: Approximately 40 minutes of data from each of 32 additional subjects to support cross-subject generalisation.

The authors provide standardized train, validation, and test splits and an open-source Python library (pnpl) that integrates with common deep learning frameworks to facilitate reproducible benchmarking. They evaluate the dataset using a word-classification task, fine-tuning the pre-trained MEG-XL model to assess performance in both within-subject and cross-subject settings.

The researchers demonstrate that scaling within-subject data is a powerful driver of decoding performance. Fine-tuning the MEG-XL model on the deep Subject 0 data significantly boosts performance for new subjects, even when those subjects contribute as little as 10 minutes of data. This suggests that pre-training on high-quality, deep neural data can effectively compensate for the practical limitations of collecting large datasets from clinical populations. The results validate the quality of the recordings and confirm that the dataset is well-positioned to serve as a benchmark for future non-invasive BCI research.