BCIwiki (bciwiki.com) — Spectral bandpower features decoded imagined speech from EEG with a mean trial-wise accuracy of 49.03% ± 4.18% in coarse-level classification across subjects, significantly higher than the 37.97% ± 3.79% achieved with time-domain statistical features, according to a preprint by Frederik Møllskov Trier, Xiaopeng Mao and Sadasivan Puthusserypady posted to arXiv on September 24, 2026. The study has not been peer reviewed.
Imagined speech decoding aims to recognize words a person silently thinks without speaking, and is seen as a potential communication route for people with severe motor impairments. The authors argue that reported performance in the field often rests on evaluation protocols that do not clearly reflect cross-subject generalization. They therefore ran a transparent baseline investigation of a multi-class imagined speech EEG dataset under a strictly subject-independent framework, testing only on subjects excluded from training.
The team compared two preprocessing and feature extraction pipelines, one based on time-domain statistics and one on frequency-domain bandpower. Both used a random forest classifier, with subject-wise cross-validation and majority voting at the trial level. Forward feature selection showed that a limited subset of frequency bands captured most of the discriminative information.
The authors say the work offers a basis for future brain-computer interface (BCI) studies seeking better cross-subject generalization. The abstract does not state the number of classes, the number of subjects or the chance level, so the practical meaning of the 49.03% figure depends on details in the full paper.