Decoding Silent Reading from Non-Invasive EEG

Researchers have made significant progress in decoding silent reading using non-invasive electroencephalography (EEG). Addressing the challenge of acquiring data for inner speech, the study used silent reading as a scalable proxy task, collecting approximately 240,000 word presentations from a single participant over 49 hours of 19-channel dry-electrode EEG. Words from continuous narrative text were presented rapidly with randomized typography to minimize visual form correlation. A convolutional EEG encoder, combined with an optional causal transformer, was trained using a CLIP-style contrastive objective to align EEG windows with word embeddings from a large language model. The results showed reliable open-vocabulary word-level decoding, extending to mid-frequency and rare words, with performance scaling log-linearly with training data volume. Control analyses confirmed that word-level decoding is distinct from narrative context tracking, establishing the feasibility of extracting rich lexical and semantic information from brain activity during silent reading.