Tap Right-⌘, speak, and your words appear at your cursor. 100% on-device dictation for macOS — free, open source, no cloud, no account.
Speech is transcribed locally by whisper.cpp (Metal GPU) with OpenAI's Whisper large-v3-turbo model. Excellent for mixed Chinese/English speech. Nothing ever leaves your machine.
- Download
VoiceTap.dmgfrom Releases and drag VoiceTap to Applications. - First launch: the app is unsigned — right-click → Open, or allow it under System Settings → Privacy & Security → "Open Anyway".
- On first run it downloads the speech model (~547MB, once). The menu-bar icon shows progress.
- Grant the permissions it asks for: Microphone (recording), Accessibility (pasting), and Input Monitoring (hotkey) if prompted. Then relaunch once.
- Focus any text field → tap Right-⌘ → speak → tap Right-⌘ → text is pasted.
- Bottom-center indicator: green bars = listening (they follow your voice), pulsing dots = transcribing.
- Recordings shorter than 0.4s are treated as accidental taps and discarded.
git clone https://github.com/2sumtech/voicetap && cd voicetap
./scripts/build-app.sh # → dist/VoiceTap.app + zip + dmg
whisper.cpp static libs for arm64 are vendored in Vendor/. To rebuild them:
git clone --depth 1 -b v1.9.1 https://github.com/ggml-org/whisper.cpp
cmake -B build -S whisper.cpp -DCMAKE_BUILD_TYPE=Release -DCMAKE_OSX_DEPLOYMENT_TARGET=13.0 \
-DBUILD_SHARED_LIBS=OFF -DGGML_METAL=ON -DGGML_METAL_EMBED_LIBRARY=ON \
-DWHISPER_BUILD_EXAMPLES=OFF -DWHISPER_BUILD_TESTS=OFF
cmake --build build -j
# copy build/src/libwhisper.a and build/ggml/src/**/libggml*.a into Vendor/lib/
- Apple Silicon, macOS 13+. Model & settings live in
~/Library/Application Support/VoiceTap/. - The only network request ever made is the one-time model download (from Hugging Face, falling back to hf-mirror.com).
MIT. Speech recognition by whisper.cpp (MIT) running OpenAI's Whisper large-v3-turbo model (MIT).