New LibriBrain100 Dataset Boosts Neural Speech Decoding Research.

Researchers have released LibriBrain100, a significant new dataset for advancing neural speech decoding. This collection comprises over 100 hours of magnetoencephalography (MEG) data, recorded while subjects listened to continuous speech. A key feature is the inclusion of approximately 80 hours from a single individual, providing an unparalleled depth of within-subject neural data, alongside additional data from 32 subjects to offer broader insights. The dataset is specifically designed for reproducible and standardized evaluation, aiming to accelerate the development of non-invasive brain-computer interfaces (BCIs). Initial evaluations using an existing decoding model on a word-classification benchmark demonstrated state-of-the-art performance, validating both the data quality and the value of extensive within-subject recordings. Furthermore, the study highlights that while deep single-subject data is powerful, broad multi-subject data can compensate for limited per-subject data through supervised finetuning of pre-trained models. The dataset comes with open-sourced tools for easy access and preprocessing, and an accompanying machine learning competition is planned to foster further research.