Subvox

joeblau/subvoxReviewed by AI

What it does

Subvox is a research prototype for macOS that tests whether the head-motion data from AirPods Pro carries enough information to turn silently mouthed words into text. You train it by holding Space and mouthing a word several times, and it then matches new movements against those recordings and shows whether the result beats chance. It is for curious technical people: there is no download, so you install xcodegen with Homebrew and build the app in Xcode. It needs macOS 14 or newer and headphones with head tracking.

How easy it is to install

E · developers only · macOS

Mac is an E: there are no releases to download, and the README's only instructions are 'brew install xcodegen', 'xcodegen generate' and building the project in Xcode. You also need AirPods Pro or other head-tracking headphones set as the Mac's audio output, and you grant Motion access on first launch.

Coming soon: install in 1 click with the RepoPicks app.

Numbers on GitHub

Stars: 337on Oct 31, 2026
Downloads: —all releases
Growth: +2,275stars per week, est.

Stars over time

No star history yet.

Facts

Runs on: macOS
You need to install: Other tools
Disk space: not determined
Accounts or keys: None
Cost: Free
License: MIT (Permissive)
Last update: Oct 31, 2026

Risk signals (5)

Attentionunproven prototype
README opening line: it is 'A macOS prototype that tests whether AirPods Pro head-motion data... carries enough information'. 'Known limits' adds that headphone motion is 'roughly 25 Hz on AirPods' and that 'Templates are user- and session-specific'.
AI review
Inforecords microphone and motion to disk
README 'Experiment protocol': recordings go to '~/Documents/Subvox/Sessions/<timestamp>-<condition>/' and include 'audio.caf: microphone audio, if the mic is running' plus every motion sample; the trained words live in '~/Library/Application Support/Subvox/dataset.json'.
AI review
Inforeads words not spoken aloud
The stated goal is to turn 'silently mouthed words into text' from headphone movement. The README says each person must record their own templates first ('Templates are user- and session-specific'), so it is not pointed at other people.
AI review
Infospecific hardware required
README 'Build & run': 'Requires macOS 14+, AirPods Pro or another headphone that supports head tracking, connected as the Mac's audio output.'
AI review
Infovery new repository
GitHub facts: created 2026-10-31 with the only commit the same day, 0 releases and 337 stars already.
AI review

Evidence is shown in its original language.

What Radar could not check

Whether it really recognises mouthed words for a new person, and how many words it can tell apart
Whether the project still builds with the current Xcode and macOS
How much disk the recorded sessions use over time
Whether anything is sent off the computer, which can only be seen by running it
How much the result changes when you take the headphones off and put them back on