Subvox Tests Silent Speech Recognition Using AirPods Motion Sensors

An open-source macOS prototype named Subvox is exploring whether consumer earbuds can detect silently mouthed words, testing the limits of standard wearable sensors against dedicated silent-speech hardware. Published on GitHub (https://github.com/joeblau/subvox) by developer Joe Blau, the application connects to Apple's AirPods Pro to determine if head-tracking motion data alone provides enough signal to distinguish unvoiced speech.

The release arrives as interest mounts around non-acoustic speech interfaces. Startup The Subvocal Company, backed by Y Combinator's Fall 2026 batch, recently announced preorders for sensor-equipped glasses that track oral movements using radio-frequency waves, according to reporting from emergent.sh. While Subvocal relies on custom hardware and foundation models to decode speech, Subvox investigates how much biomechanical signal is already accessible through consumer devices without specialized hardware.

Subvox runs on macOS 14 and queries Apple's CMHeadphoneMotionManager framework for motion telemetry. Rather than employing large neural networks or language models, the software functions as a constrained, small-vocabulary template recognizer. When a user silently mouths a word, Subvox records nine motion dimensions per frame: three axes each of user acceleration, rotation rate, and attitude displacement. Each channel is mean-centered, and quiet lead-in and lead-out frames are trimmed. The software scales the channels and runs a one-nearest-neighbor match using dynamic time warping with a Sakoe–Chiba band, rejecting classifications if candidate distances exceed thresholds or fail confidence margins.

To establish whether the recognition reflects a genuine signal rather than random motion artifacts, the software includes a statistical validation suite. Subvox pairs leave-one-out cross-validation with a 500-iteration label-permutation test, requiring a p-value below 0.05 to confirm that the motion stream separates words more reliably than chance.

The project also documents notable constraints inherent to commercial audio hardware. Blau notes that Apple's headphone motion feed is processed and updates at a relatively low rate of approximately 25 Hz. Furthermore, Apple does not expose raw data from its internal speech-detecting accelerometers or inward-facing microphones to third-party software. Consequently, the tool captures gross jaw articulation and secondary head movement rather than direct vocal-tract acoustic vibrations. The resulting templates remain user- and session-specific, and re-seating the AirPods can alter the kinetic baseline. Selecting AirPods as the active microphone input also forces the earbuds into the lower-bandwidth Bluetooth Hands-Free Profile.

The prototype highlights the trade-off between commodity peripherals and bespoke hardware in silent computing. As startups like Subvocal work toward delivering dedicated RF-sensing wearables in early 2027, Subvox provides an open benchmark for how far low-frequency inertial data can go before specialized sensors become mandatory.