Silent dictation for places where speaking aloud would interrupt everyone else: Aarush Yadav and Arnav’s Subvocal prototype uses radio signals from glasses to detect mouth movements, then a trained model decodes them into text.

The makers describe moving from facial electrodes to radio sensing and collecting calibration data across people. This is deliberate silent mouthing, not thought reading.

Their original demonstration shows a partner trying the prototype. Outside viewers appreciate its candid presentation, but it is not an independent accuracy benchmark. The planned consumer product has not shipped; decoding currently uses a companion app and cloud service.

Maker’s technical account and discussion: news.ycombinator.com/item?id=50340088