WIP Special
Oiwi Parker Jones
Wednesday, 26 November 2025, 12pm to 1pm
Hybrid via Teams or in-person in the Cowey rooms, FMRIB Annexe
Hosted by Camille Lasbareilles
Join via Teams
LibriBrain100
Preseted by Oiwi Parker Jones
Abstract: The aim of this study is to acquire a "large" and "naturalistic" dataset of speech processing for the development of neural decoding methods. By "large" and "naturalistic", the dataset will include over 100 hours of MEG recordings acquired while subjects listen to audiobooks. This will build on the LibriBrain dataset (Özdogan et al. 2025) which we published at NeurIPS and has formed the basis for the 2025 PNPL Competition (Landau et al. 2025) - also published at NeurIPS. Both datasets significantly advance recent work in large and naturalistic speech datasets (e.g. Armeni et al. 2022; Gwilliams et al. 2023; d'Ascoli et al. 2024). There are two main differences between LibriBrain and LibriBrain100. First, LibriBrain100 will double the amount of data (from 50 hours to 100 hours). Second, whereas LibriBrain focused on so-called "deep" (i.e. large-scale within subject) data, LibriBrain100 aims both for "depth" and "bredth" - extending the number of subjects significantly (64 subjects), allowing the investigation of individual variation in speech processing and between-subject transfer learning. The plan is for these data to form the basis of another open machine learning competition (cf. https://libribrain.com/).