LibriBrain100 Preprint Introduces Largest Within-Subject MEG Dataset for Speech Decoding

Summary

Researchers led by Francesco Mantegna, with senior author Oiwi Parker Jones, released LibriBrain100, a dataset more than doubling the size of the original LibriBrain release to over 100 hours of high-quality MEG recorded while subjects listened to naturalistic continuous speech. With roughly 80 hours from a single subject, the dataset sets a new record for within-subject depth in non-invasive neural speech data -- eight times deeper than the next comparable dataset. Using an existing decoding model on a word-classification benchmark, the team achieved state-of-the-art performance, validating both the recording quality and the value of large within-subject datasets.