All Quiet on the Frontal Lobe: Physiological Noise \\ Augmentation for Non-Invasive Brain-to-Speech

Non-invasive brain-to-speech decoding aims to restore communication to patients suffering from neurodegenerative disease, without the risks of neurosurgery. Existing M/EEG-based methods, while scalable, continue to suffer from high word error rates driven by extremely low signal-to-noise ratios as task-agnostic brain activity dominates recordings. We propose \textit{physiological noise augmentation} (PNA), an ICA-based data augmentation method that explicitly trains decoders to become invariant to artifacts (e.g. ocular and cardiac activity), using NLP-inspired feature remixing to generate biophysically realistic, label-preserving training examples. Combining PNA with multi-trial averaging suppresses residual unmodeled variability that is not phase-locked across repeated trials. We further show that, to first order, PNA approximates anisotropic regularization that penalizes decoder sensitivity along artifact-dominated directions. On MegNIST, a 12-hour imagined-digit MEG dataset, PNA with 10-trial averaging achieves $70.9\%$ decoding accuracy using EEGNet, improving performance by $8.76\%$ over training on real data alone. Our results suggest that artifact-aware augmentation and trial averaging are complementary tools for improving robustness in non-invasive speech BCIs.