Computer Science

Decoding speech perception from non-invasive brain recordings

A. Défossez, C. Caucheteux, et al.

Using contrastive learning on non-invasive MEG and EEG recordings from 175 volunteers, this study decodes perceived speech representations—identifying, from 3 seconds of MEG, the correct speech segment among 1,000+ candidates with up to 41% average accuracy and 80% for top participants. The research was conducted by Alexandre Défossez, Charlotte Caucheteux, Jérémy Rapin, Ori Kabeli and Jean-Rémi King.