BCIwiki (bciwiki.com) — According to arXiv, Blessing Isoyiza Adeika and Onyema Osuagwu uploaded a preprint on cross-subject imagined speech decoding on October 14, 2026. It has not been peer reviewed. The study tested whether imagined speech content recorded with scalp EEG transfers across people, examining a requirement for a non-invasive brain-computer interface (BCI) that would need no calibration for a new user, rather than reporting an achieved speech interface.
The researchers analyzed three public corpora: ds003626, KaraOne, and Track 3 of the 2020 International Brain-Computer Interface Competition. Their protocol enforced split integrity through assertions and reported exact binomial thresholds, group tests, and bootstrap confidence intervals. Decoders were trained on each participant and tested on every other participant. The team also compared representational geometry across people without a classifier, applied subject-level alignment, and tested five explanations for limited within-subject performance.
Across 482 ordered participant pairs, none remained significant after multiple-comparison correction. Mean pair-transfer accuracy was within 0.006 of chance in every corpus, while every within-subject reference was above chance. The authors concluded that imagined speech content did not transfer across people in these corpora through any of the representations they examined.
Imagery versus pre-stimulus rest generalized across people at an accuracy of about 0.61 after controls for artifacts and elapsed time. However, the decoder labeled the interval preceding imagery as imagery as often as it labeled imagery itself. The study therefore interpreted this contrast as reflecting a post-stimulus state rather than an imagery-specific signal.
Within subjects, only the competition corpus exceeded the decoding threshold for every participant, with a reported accuracy of 0.3147 against a chance level of 0.200. Syllable count explained little of this result, and its frequency profile failed an artifact audit. The authors identified trial count per class as the only variable that ordered the corpora and favored larger single-condition datasets over further model development. This is their interpretation, not causal evidence that increasing trial counts necessarily improves decoding.