BCIwiki (bciwiki.com) — According to arXiv, Shahryar G. Namdari, Francesco Iacomi, Riccardo Barbieri, and Luca Mainardi uploaded a preprint on imagined speech decoding to the platform on October 14, 2026. It has not been peer reviewed. The study aligns EEG with text representations to investigate whether a non-invasive brain-computer interface (BCI) can preserve semantic relationships while recognizing fixed intents.
Traditional closed-set classifiers select from predefined categories and typically treat semantic categories as independent labels. The researchers adapted multimodal contrastive learning to map EEG trials into the embedding space of a frozen pretrained text encoder, a numerical space for comparing language representations. Training combined semantically reweighted EEG-sentence alignment with text-prototype-based class discrimination to encourage semantic structure while maintaining separability among 39 intent categories.
The framework was evaluated on five subjects in the Chisco imagined speech dataset using Top-k accuracy, macro-F1, and balanced pairwise matching accuracy (2v2). Its decoding performance was comparable to that of a task-optimized closed-set classifier. Maximum EEG-text cosine similarity (MaxCosine), investigated as an intrinsic reliability indicator, identified held-out trial subsets with higher Top-1 and Top-5 accuracy across subjects. This finding concerns subset selection, not an accuracy increase across all trials.
Five evaluation-time semantic axes and a qualitative Semantic Gallery Probe (SGP) examined semantic structure beyond category labels. Mean agreement exceeded the 50% chance reference on all five dimensions, while the SGP illustrated direct comparisons with previously unseen candidate sentences. The authors interpret these findings as suggesting that EEG-language alignment can retain fixed-intent decoding performance while supporting similarity-based retrieval, intrinsic reliability ranking, and broader semantic evaluation. The qualitative demonstrations do not establish unrestricted decoding of arbitrary imagined sentences.