Decoding imagined speech from EEG offers a potential communication pathway for individuals with severe motor impairments, yet many reported results do not clearly reflect cross‑subject generalization. This paper conducts a transparent baseline study on a multi‑class imagined speech EEG dataset under a strictly subject‑independent evaluation framework.
Two preprocessing and feature extraction pipelines are compared: a time‑domain statistical feature approach and a frequency‑domain spectral band‑power approach. Both pipelines use a random forest classifier, evaluated with subject‑wise cross‑validation and trial‑level majority voting.
The spectral pipeline achieved a mean trial‑wise accuracy of $49.03 \pm 4.18\%$, significantly higher than the statistical pipeline’s $37.97 \pm 3.79\%$. Forward feature selection further indicated that a limited subset of frequency bands captured most of the discriminative information.
Overall, this work provides a clear baseline and a strong foundation for future brain‑computer interface studies aiming to improve cross‑subject generalization in EEG‑based imagined speech decoding.
Review: By adopting a rigorous subject‑independent protocol, the study highlights the superiority of frequency‑domain features and points toward more generalizable BCI designs.