XSci

MEGWeaver: Sentence Relational Pretraining for Audio-Free MEG-to-Text Decoding

Chenyuan Zhu, Huanran Zheng, Dell Zhang, Xiaoling Wang

Submitted by Chenyuan Zhu

Published November 17, 2026 · Version v1, November 17, 2026 · DOI 10.66977/xsci.2611.002u

Artificial Intelligence, Machine Learning, Biophysics

Abstract

Decoding perceived language from non-invasive magnetoencephalography (MEG) offers a promising route toward brain-to-text communication. NeuSpeech first introduces Whisper to MEG-to-text decoding, while BrainECHO improves its accuracy through acoustic reconstruction. However, whether further accuracy gains can be achieved without relying on additional acoustic supervision remains unclear. Here, we propose MEGWeaver, a framework for direct MEG-to-text generation that pretrains a MEG encoder based on sentence relational pretraining (SRP), which aligns the in-batch similarity structure of MEG representations with sentence-level relations defined by similarities between embeddings from a frozen sentence encoder. Only the pretrained sensor projection is transferred to downstream text decoding. Across English, Dutch and Chinese datasets under held-out-participant evaluation, MEGWeaver achieves BLEU-4 scores of 75.10±1.2175.10\pm1.21, 48.78±1.1748.78\pm1.17 and 39.07±0.7239.07\pm0.72, respectively, outperforming NeuSpeech and sentence retrieval across all three datasets and obtaining numerically higher BLEU-4 than our BrainECHO reproduction on Dutch and Chinese. These findings support direct MEG-to-text generation for known sentences under the present segmentation protocol, while decoding novel sentences remains an open problem.

View PDF