Abstract Background: Over 20 million people worldwide lose speech following stroke or Parkinson's disease, yet many cannot access current brain-computer interfaces due to surgical risks or contact sensitivities. 
Objective: We demonstrate contactless optical decoding of binary inner speech ("yes" versus "no") using laser speckle-pattern analysis.
Methods: Speckle-pattern dynamics were recorded at 1000 fps from the scalp overlying Broca's area during silent inner speech in 10 healthy volunteers. Deep learning models were trained on millions of video frames using a self-supervised long-video masked autoencoder (LV-MAE) for representation learning, followed by lightweight classifier adaptation requiring only one minute of subject-specific calibration per class.
Results: Classifiers achieved a mean AUC of 0.97 and an accuracy of 95.7% on 40-ms inputs (10-fold cross-validation, 3,180 s of balanced recordings). LV-MAE representations proved effective for speckle-based cortical decoding. Forehead controls showed below-chance classification, supporting the cortical origin of the signal. Model rank ordering remained stable across recordings obtained on the same day for 10 participants. In the one subject retested after one month, 1s decoding showed a slight decline in accuracy (98.53% → 87.1%, after threshold recalibration using 1s of data per class) while AUC remained unchanged. 
Conclusions: These findings demonstrate a proof of concept for a contactless binary inner speech decoding in healthy volunteers. Translation to real-world BCI applications will require clinical validation in patient populations and an extension beyond binary vocabulary.