Description
Abstract—Electroencephalography (EEG) brain–computer in-
terfaces (BCIs) enable communication without muscular output,
yet most BCI typing systems rely on decision-complete neural
events (e.g., evoked potentials) that impose stimulus-driven se-
lection loops and seconds-scale latency. This paper studies a
different operating point: predicting a user’s intended next word
during early stages of language formulation, before a conscious
lexical commitment is complete. We propose an EEG-based
typing assistant that leverages weak, pre-commitment signals
by decoding in a frozen language-model embedding space and
fusing with a transformer language model through conservative,
uncertainty-aware constraints. Our method introduces (i) Pre-
Form, a temporal EEG encoder that predicts distributions over
semantic token embeddings; (ii) CNTA, a contrastive neuro-token
alignment objective that improves early-window decoding and
mitigates prompt memorization; (iii) Neuro-Gated LM, a gating-
and-masking fusion mechanism with a reject option that pre-
vents language-model dominance; and (iv) ASK, a meta-learned,
parameter-efficient personalization scheme enabling minutes-
scale adaptation to new users. We evaluate under a reproducible
cued silent word-planning protocol with leave-one-subject-out
testing and ablations, reporting top-k accuracy, word error rate
(WER), latency-to-correct-suggestion, information transfer rate
(ITR), and expected calibration error (ECE). Offline results
suggest that conservative next-word suggestions can be surfaced
hundreds of milliseconds earlier than decision-complete baselines
while maintaining calibrated abstention behavior.
Reviews
There are no reviews yet.