RAG-based EEG-to-Text Translation Using Deep Learning and LLMs
A new study proposes a retrieval-augmented generation (RAG)-based method for translating EEG signals into text. This approach aims to improve sentence-level decoding, which has been challenging due to low signal-to-noise ratios in EEG recordings. The proposed pipeline shows a significant improvement over random baseline performance in experiments conducted on EEG data from silent reading.
- ▪The study introduces a RAG-based EEG-to-text decoding pipeline that combines an EEG encoder with semantic sentence embeddings.
- ▪Experiments were conducted using the Zurich Cognitive Language Processing Corpus, which includes single-trial EEG recordings.
- ▪The proposed method achieved a mean cosine similarity of 0.181, representing a 30.45% improvement over the random baseline.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.17503 (cs) [Submitted on 17 May 2026] Title:RAG-based EEG-to-Text Translation Using Deep Learning and LLMs Authors:Enrico Collautti, Xiaopeng Mao, Luca Tonin, Stefano Tortora, Sadasivan Puthusserypady View a PDF of the paper titled RAG-based EEG-to-Text Translation Using Deep Learning and LLMs, by Enrico Collautti and 4 other authors View PDF HTML (experimental) Abstract:The decoding of linguistic information from electroencephalography (EEG) signals remains an extremely challenging problem in brain-computer interface (BCI) research. In particular, sentence-level decoding from EEG is difficult due to the low signal-to-noise ratio of these recordings.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
