Postprocessing of P300 Speller Output with a Large Language Model

Abstract

P300 spellers convert electroencephalographic (EEG) activity into text by presenting users with a matrix of flickering characters. While these systems can achieve high classification accuracy, communication is severely slowed by the need for many stimulus repetitions to obtain a reliable signal. Reducing repetitions accelerates spelling but introduces character-level errors: insertions, deletions, and substitutions that degrade usability and increase user fatigue. Although substantial research has focused on improving performance at the signal acquisition and decoding stages, here we investigate a complementary text post-processing approach that leverages large language models (LLMs) to restore corrupted P300 speller output. We constructed a dataset derived from cLang-8 and simulated realistic P300-style text corruption using both random and empirically derived human-like error strategies. We evaluated several instruction-tuned LLMs alongside an optical character recognition (OCR)-fine-tuned ByT5 model under zero-shot and few-shot prompting conditions. We found that LLMs effectively recovered clean text from noisy inputs. Models employing SentencePiece tokenization consistently outperformed byte-pair encoding (BPE)-based counterparts, and few-shot in-context learning further improved restoration accuracy, with Gemma 3 achieving the strongest performance across all settings. These results suggest that LLM-based post-processing could enable P300 speller systems to operate with fewer repetitions and lower latency while maintaining or improving output accuracy, offering a practical path toward more efficient daily communication for users with motor and speech disabilities.