BCIwiki (bciwiki.com) — A team at the Research Center of Neurotechnologies, Southern Federal University, Russia, created a directional-word EEG dataset, published in Scientific Data on August 17, 2026. It records overt and covert articulation in 22 neurologically healthy volunteers for research and benchmarking of inner-speech BCI algorithms.
The researchers recorded EEG from 22 neurologically healthy volunteers during overt and covert articulation of six spatial-direction words. The group comprised 12 native Russian speakers and 10 native Spanish speakers. The Russian protocol also included an additional “next” command. Monopolar EEG signals were acquired using 38 electrodes positioned according to the international 10-10 system and a Neurovisor-BMM-52 (NVX) amplifier at a sampling rate of 500 Hz.
Six participants had concurrent EMG from the masseter and laryngeal region: two Russian and four Spanish speakers. Exploratory classification used power and coherence features in the 55–65 Hz band with per-participant five-fold cross-validation. Random Forest reached 78 ± 4% accuracy for distinguishing rest, overt speech and inner speech; SVM reached 74 ± 5% and LDA 67 ± 6%. These figures are not six-word recognition accuracies.
The authors describe classification as an exploratory data-quality check. Keypress markers, articulatory muscle activity and largely fixed word order may introduce confounds; some recordings also used different marker protocols. The dataset does not establish unrestricted inner-speech decoding, improved patient communication or cross-language decoding.