A public GitHub repository by Joe Blau describes Subvox, a macOS prototype that tests whether AirPods Pro head-motion data can turn silently mouthed words into text¹.

The README says the project does not use a language model¹. It is a small-vocabulary template recognizer plus a statistical test meant to show whether the motion stream distinguishes words better than chance¹.

The app requires macOS 14 or later and AirPods Pro or another headphone that supports head tracking, connected as the Mac's audio output¹. It reads motion through Apple's CMHeadphoneMotionManager API¹. On first launch, the project asks for Motion permission¹.

A Sensors screen plots live acceleration, rotation rate and attitude, sample rate, and an optional microphone meter¹. On Train, a user mouths a chosen word while holding the space bar; each take is stored as a template¹. Evaluate runs leave-one-out classification and a 500-repeat label-permutation test, then reports accuracy against chance, a p-value and a confusion matrix¹.

Decode matches a new segment to templates with dynamic time warping and can reject a guess if the distance is too large or the margin is too small¹. A live microphone transcript can run alongside for comparison¹.

The README lists limits: headphone motion is processed and roughly 25 hertz, and Apple does not expose a speech-detecting accelerometer or inward microphone, so the system is expected to see jaw and head movement rather than vocal vibration¹. Templates are user- and session-specific, and reseating the earbuds can change the signal¹.

The repository is MIT-licensed, created Oct. 31, 2026, and includes an Xcode project generated with xcodegen¹. Independent tests of accuracy outside the author's own protocol are not in the README¹.

Sources

1. github.comSubvocal raised $1.3M. A developer recreated it with AirPods in 9 minutes3:40 PM

Read the original at github.com ↗