Neural Spelling: A Spell-Based BCI System for Language Neural Decoding

Brain-computer interfaces (BCIs) support the study of communication-oriented neural decoding by translating neural activity into text, yet existing non-invasive systems rarely cover the full alphabet in handwriting-based settings.

We propose a novel non-invasive EEG-based BCI framework, Curriculum-based Neural Spelling (CNS), that decodes all 26 English letters by first learning neural patterns associated with handwriting trajectories. A Generative AI (GenAI) module based on large language models (LLMs) is then integrated to transform noisy letter-level neural predictions into sentence-level outputs under explicit neural constraints.

The proposed system achieves robust letter-level decoding and improved sentence-level reconstruction under controlled offline evaluation, outperforming conventional EEGNet and hybrid CNN-RNN baselines. GenAI correction further reduces word error rates and enhances decoding fluency.

Combining EEG-based neural spelling with generative language modeling supports the study of full-alphabet decoding and improves sentence-level linguistic metrics in a controlled non-invasive EEG setting, but does not by itself establish clinical or real-world usability.

This work demonstrates how integrating GenAI with neural decoding can bridge the gap between noisy signal-level predictions and coherent language-level outputs, establishing a system-level framework for full-alphabet neural spelling and adaptive language-level correction under non-invasive EEG constraints.