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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Non-invasive brain-computer interfaces (BCIs) are crucial for communication-oriented neural decoding.
• Existing systems often struggle to decode the full alphabet, particularly in handwriting-based tasks.

Purpose of the Study:

• To develop a novel non-invasive EEG-based BCI framework capable of decoding all 26 English letters.
• To integrate Generative AI (GenAI) for improving the accuracy and fluency of neural decoding outputs.

Main Methods:

• Proposed a Curriculum-based Neural Spelling (CNS) framework using electroencephalography (EEG).
• Trained the system to recognize neural patterns associated with handwriting trajectories for letter decoding.
• Integrated a Generative AI (GenAI) module, leveraging large language models (LLMs), to refine letter-level predictions into coherent sentences.

Main Results:

• Achieved robust letter-level decoding accuracy in offline evaluations.
• Demonstrated improved sentence-level reconstruction compared to conventional EEGNet and CNN-RNN models.
• GenAI-powered correction significantly reduced word error rates and enhanced decoding fluency.

Conclusions:

• The integration of EEG-based neural spelling with generative language modeling enables full-alphabet decoding.
• This framework enhances sentence-level linguistic metrics within a controlled non-invasive EEG setting.
• While promising, the system's clinical or real-world usability requires further investigation.