Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain
The release gives speech-decoding researchers 30 hours of whole-brain 3T fMRI per subject, tied to podcast audio and transcripts.
Three healthy subjects were scanned across multiple sessions while listening to native-language podcasts. The dataset includes aligned brain activity, audio, transcript annotations, and derived stimulus embeddings. The authors also define a benchmark where a decoder must retrieve the matching podcast segment from held-out fMRI activity. Their scaling analysis reports that decoding improves as more per-subject training data is used. ArXiv · AI/CL/LG's note
Three healthy subjects were scanned across multiple sessions while listening to native-language podcasts. The dataset includes aligned brain activity, audio, transcript annotations, and derived stimulus embeddings. The authors also define a benchmark where a decoder must retrieve the matching podcast segment from held-out fMRI activity. Their scaling analysis reports that decoding improves as more per-subject training data is used. ArXiv · AI/CL/LG's note
score 5