Reconstructing the brain’s language network from non-invasive EEG enables interpretable speech decoding

<title>Abstract</title> <p>Non-invasive EEG-based speech decoding typically treats the neural-to-text mapping as a black box, offering limited interpretability and few explicit links to the brain’s language networks. Here we present a framework that reconstructs language-network dynamics in source space from scalp EEG by integrating neural perturbational inference with geometric constraints derived from cortical eigenmodes. A meta-learning diffusion decoder then translates these dynamics into text with few-shot adaptation. Evaluated on Chinese and English reading EEG, the framework achieves state-of-the-art decoding performance, reducing the character error rate by up to 11% relative to existing methods. The decoded semantic fields localize to temporal regions that match the ventral language pathway, and cross-modal experiments reveal shared semantic representations between reading and listening. Moreover, orthographic, omission, and semantic errors map onto distinct electrophysiological signatures—the N170, P200, and N400 components, respectively—and error-informed correction using these signatures further reduces the character error rate by 8.3%. By establishing an interpretable signal-to-language-network-to-text pathway, this work advances non-invasive brain–computer interfaces toward physiologically grounded, cross-modally generalizable, and diagnostically meaningful communication</p>