Remote Optical Decoding of Inner Speech in Broca’s Area via AI-based Speckle Pattern Analysis

Abstract Over 20 million people worldwide with stroke or Parkinson's disease experience speech loss, yet surgical risks or contact sensitivities exclude many from current brain-computer interfaces (BCIs). Here, we report a contactless approach for decoding binary inner speech ("yes" versus "no") from laser speckle-pattern dynamics recorded over Broca's area. Using deep learning on millions of speckle-pattern video frames from 10 healthy volunteers, classifiers achieved a mean AUC of 0.97 and an accuracy of 95.7% for 40-ms inputs (10-fold cross-validation on 3,180 seconds of balanced recordings), with minimal calibration (one minute per subject). This work demonstrates that representations learned with a long-video masked autoencoder (LV-MAE) are effective for speckle-based cortical brain imaging, establishing the feasibility of applying self-supervised long-video architectures to this modality. The forehead control recordings yielded below-chance separability on the control validation set, consistent with the absence of a stable discriminative signal from the control region and with a contribution from cortical signals. These results demonstrate proof of concept for contactless binary inner speech decoding in healthy volunteers. Translation to real-world BCI applications will require clinical validation in patient populations and extension beyond binary vocabulary.