Paper: non-invasive brain-to-text gains were a timing leak

A neural network given synthetic signals containing no brain information decoded perceived speech at 22.0% balanced accuracy. The same architecture, fed actual brain recordings, reached 22.3%.

That near-identical split is the subject of a new arxiv preprint by Dulhan Jayalath and Oiwi Parker Jones, who argue the influential d'Ascoli et al. (2025) brain-to-text result leaked word timing into its predictions. In the method under scrutiny, 'time series of brain activity from subjects perceiving continuous speech are segmented into fixed-length windows starting at each word.' A single network then predicts every word in a sentence jointly. Neighbouring windows overlap, 'implicitly revealing the interval between words.'

Those intervals carry word durations. 'The' is much shorter than 'supercalifragilisticexpialidocious.' The network could pick up word length without touching neural signal.

The fix is one line of pipeline logic: 'Instead of jointly encoding all windows in a sentence, we process each independently.' With that change, two older ideas start working again: aggregating predictions across repeated neural responses to the same word, and using a pretrained LLM as a linguistic prior. Together they produce SimpleB2T, which reports 'a word error rate of 36.6% with five observations per word, approaching past invasive speech decoding performance, albeit under different conditions.'

The abstract names no other papers as affected. Jayalath and Parker Jones frame their contribution as exposing 'an important shortcut in brain-to-text decoding,' and put no number on how much of the broader non-invasive progress reported since d'Ascoli et al. rests on the same window design.