Cross-subject decoding of human neural data for speech Brain Computer Interfaces.

Boccato, Tommaso; Olak, Michal; Ferrante, Matteo · J Neural Eng · 2026

basic_science · Level V

Where this comes from

Abstract

<b>Objective</b>: Brain-to-text systems have recently achieved impressive performance when trained on single-participant data, &#xD;but remain limited by uninvestigated cross-subject generalization. &#xD;&#xD;<b>Approach:</b>We present the first neural-to-phoneme decoder trained jointly on the two largest intracortical speech datasets &#xD;(Willett et al. 2023; Card et al. 2024), &#xD;introducing day- and dataset-specific affine transforms to align neural activity into a shared space. &#xD;Additionally, a hierarchical GRU decoder with intermediate CTC supervision and feedback connections is designed to address&#xD;the conditional-independence assumption of standard CTC loss. &#xD;&#xD;<b>Main Results:</b>Our model matches or outperforms within-subject baselines while being trained across participants, &#xD;and adapts to unseen subjects using only a linear transform or brief fine-tuning. &#xD;On an independent inner-speech dataset (Kunz et al. 2025), &#xD;our approach shows some initial evidence of generalization, by training only subject-, day-specific transforms. &#xD;&#xD;<b>Significance:</b>These results demonstrate the feasibility of cross-subject pretraining as a promising direction toward more scalable speech BCIs.