BCIwiki (bciwiki.com) — According to arXiv, Blessing Isoyiza Adeika and Onyema Osuagwu uploaded a preprint auditing scalp-EEG speech decoding on October 14, 2026. It has not been peer reviewed. The study examines whether overt speech can validate a decoding pipeline as a positive control and what detection capability underlies a null result. A positive control is a reference condition expected to yield a detectable result and used to check a pipeline.
The researchers analyzed ds003626 ("Thinking out loud"), KaraOne, and Track 3 of the 2020 International Brain-Computer Interface Competition, auditing overt speech, articulator preparation, and imagined speech. Their protocol enforced split integrity through assertions and used artifact-channel subsets, band-limited decoding, and trial selection. They also injected class-dependent effects of known amplitude into real recordings to measure detection floors for five decoder families and checked ds003626 labels against its acquisition protocol.
On ds003626, two lip electrodes decoded overt speech with an accuracy of 0.7430, and the eight external channels achieved 0.7068, compared with 0.4231 for the 128-channel scalp montage. No decodable signal survived restriction to the 8-30 Hz band. In the other two corpora, within-subject accuracy increased monotonically with frequency into the range dominated by muscle activity. The authors interpret these findings as showing that myogenic signals dominate the overt-speech positive controls in these corpora, preventing those controls from validating a pipeline's ability to decode brain-derived speech signals.
Under a criterion fixed in advance, band-power logistic regression detected an injected 8-30 Hz desynchronization of 0.54 uV. By contrast, the compact convolutional network EEGNet detected no effect within subjects at injected amplitudes up to 4.31 uV. These values describe detection tests using artificially injected effects, not measurements of natural speech-related brain-signal amplitudes or evidence that one model is superior across all tasks.
The label audit found that ds003626 condition codes reverse the intuitive reading, class encoding changed between public releases, and condition is a deterministic function of time within each session. The authors recommend reporting band-limited and artifact-channel accuracy, the direction of accuracy changes across frequency, and a measured detection floor.