Free, local, on-device voice input for any app on your Mac.
Start the service once. Then switch to Claude, Cursor, VS Code, your browser, or any other window — hold Fn+F10, speak, release. Your words appear as text, instantly, in whatever app is focused. No cloud, no API key, no subscription.
| LocalVoiceAI | Cloud STT (Whisper API, etc.) | |
|---|---|---|
| Cost | Free | Pay per minute |
| Privacy | Fully on-device | Audio sent to cloud |
| Latency | ~1–2s (Metal GPU) | Network dependent |
| Works offline | Yes | No |
| Setup | One-time | API key required |
- Apple Silicon Mac (M1 or later)
- macOS 14+
- Homebrew
1. Download and install
curl -sL https://github.com/npateriya/LocalVoiceAI/releases/latest/download/localvoiceai-darwin-arm64.tar.gz | tar -xz
cd localvoiceai-darwin-arm64
bash install.sh1. Install dependencies and build
make setup # install Homebrew dependencies
make install # build and install2. Grant permissions (one-time — persists across reboots)
In System Settings → Privacy & Security, add ~/.local/bin/localvoice to both:
Accessibility — allows localvoice to simulate Cmd+V paste into the active window
Input Monitoring — allows localvoice to detect the F10 keypress system-wide
Both toggles must be blue (enabled). After
make update(rebuild), re-grant both — the binary hash changes.
3. Start the service
launchctl start com.localvoiceai.localvoiceOr if installed from source:
make start4. Use it
Switch to any app — Claude, Cursor, VS Code, browser, anything — hold Fn+F10, speak, release. Text appears at the cursor.
Check logs:
tail -f /tmp/localvoice.log- Key tap — a C-level event tap on a dedicated thread watches for F10 key down/up
- Record — PortAudio captures mic audio at the device's native sample rate (48kHz / 44.1kHz)
- Transcribe —
whisper-cliruns the Whisper model locally on Apple Metal GPU (~1–2s for a short phrase) - Paste — text is copied to clipboard and Cmd+V is simulated into the focused window
Everything runs on your machine. No data ever leaves it.
LocalVoiceAI runs as a macOS LaunchAgent — independent of any terminal or IDE. This means it works consistently whether you launched it from Terminal, iTerm2, VS Code's terminal, or anywhere else.
| Command | What it does | Re-grant permissions? |
|---|---|---|
make install |
First-time setup | Yes (once) |
make start |
Start listening | No |
make stop |
Stop listening | No |
make reload |
Reload config (no binary change) | No |
make update |
Rebuild + reinstall binary | Yes |
make uninstall |
Remove everything | — |
Override the push-to-talk key:
WHISPER_KEYCODE=101 make start # use F9 instead of F10Common keycodes:
| Key | Code |
|---|---|
| F8 | 100 |
| F9 | 101 |
| F10 | 109 (default) |
| F11 | 103 |
| F12 | 111 |
The Whisper ggml-small model (~244MB) is downloaded automatically to ~/.cache/localvoice/ on first run. For different accuracy/speed tradeoffs, download another model from HuggingFace ggerganov/whisper.cpp and place it in ~/.cache/localvoice/.
Event tap failed — One or both permissions are missing. Re-add ~/.local/bin/localvoice to Accessibility and Input Monitoring in System Settings.
Doesn't work in iTerm2 / VS Code terminal — Use make install + make start instead of running ./localvoice directly. The LaunchAgent runs as its own process, bypassing terminal permission restrictions.
Transcription is gibberish — Usually a sample rate mismatch. localvoice auto-detects the native mic rate; no config needed.
Permissions lost after make update — The binary hash changed. Re-add ~/.local/bin/localvoice to both permission lists after rebuilding.
- Key monitoring —
CGEventTap(kCGSessionEventTap, listen-only) on a C pthread, piped to Go to avoid CFRunLoop/goroutine conflicts - Audio capture — PortAudio at the mic's native sample rate
- Transcription —
whisper-cli(Homebrewwhisper-cpp) with Metal GPU acceleration - Paste —
pbcopy+CGEventPostsimulating Cmd+V - Service — macOS LaunchAgent for terminal-independent, persistent operation
Apache License 2.0 — see LICENSE.