Subvocal raised $1.3M. A developer recreated it with AirPods in 9 minutes

Subvox is an experimental macOS app investigating whether AirPods Pro head-motion sensors can distinguish words silently mouthed by a user and convert them into text. It uses no language model: users create templates by mouthing selected words, then test or decode new segments. The app records acceleration, rotation, and changes in headphone orientation, and can capture microphone audio for comparison. Segments are started by holding Space or by an automatic trigger calibrated to background noise. Suggested experiments compare normal speech, whispering, and silent mouthing across repeated attempts.

For recognition, Subvox normalizes the sensor channels and matches each new segment to the closest trained word template using dynamic time warping. It can reject uncertain matches based on distance and the gap between the best and second-best choices. Its evaluation uses leave-one-out classification and compares accuracy with results from 500 shuffled-label tests, helping determine whether the motion signal carries more information than chance. The project is a feasibility test, not a demonstrated general-purpose speech system: AirPods motion data is processed at roughly 25 samples per second, and the relevant speech-detection accelerometer and inward-facing microphone are unavailable. The signal is therefore expected to reflect jaw and head movement rather than vocal vibration. Templates may vary by user, session, and how the earbuds sit, limiting reliability.