Free, private voice dictation for macOS. Hold Fn, speak, let go, and clean text lands in whatever app you were typing in.
DaemonFlow is an on-device alternative to Wispr Flow. It transcribes with Whisper or NVIDIA Parakeet, then a small local language model strips the ums, applies your self-corrections and fixes punctuation. Your voice and your text never leave your Mac.
- Dictate anywhere: hold Fn to talk and release to paste into the frontmost app. Double-tap Fn for hands-free mode, and tap it again to stop.
- Live preview: a small pill above the Dock shows what it's hearing while you talk.
- On-device transcription: Whisper (Tiny to Large v3 Turbo) through whisper.cpp with Metal, or Parakeet TDT 0.6B v3 on the Neural Engine.
- Local cleanup: an MLX language model (Llama 3.2, Qwen 2.5, Phi 3.5 or Llama 3.1 8B) turns speech into written text: filler words removed, "Thursday, no wait, Friday" resolved to "Friday", and spoken lists formatted as lists.
- Screen awareness: names and jargon in the window you're working in become spelling hints, so "Supabase" on screen stops coming out as "super base".
- Custom vocabulary: add the spellings you keep fixing by hand, and anything close enough is rewritten to match.
- ⌥1 refine: select text in any app and tap ⌥1 to clean it up in place. Hold ⌥1 and say how to change it ("make it more formal").
- Editable prompts: the instructions behind dictation cleanup, refine and instructed refine are yours to change.
- Your pill, your way: set the pill's size, colours and border.
- History: your last 200 dictations, raw and cleaned, stored on your Mac.
When you press Fn, DaemonFlow reads the text of the frontmost window through the Accessibility API and keeps only the unusual words: product names, identifiers and anything the macOS spell checker doesn't know. Those terms are passed to Whisper as a prompt and to the cleanup model as spelling references. Password fields are never read. Turn it off under Vocabulary → Use text on screen.
- macOS 27 or later
- Apple Silicon (the cleanup models run on MLX)
- Download
DaemonFlow-macOS.zipfrom Releases, unzip it and drag DaemonFlow to Applications. - The app isn't notarized yet, so open it the first time with right-click → Open → Open, or run
xattr -cr /Applications/DaemonFlow.apponce. - Grant Accessibility (to paste and read the screen), Input Monitoring (for the Fn key) and Microphone when the setup asks.
- Pick a transcription model and a cleanup model. Each downloads once, then works offline.
- Click into a text field in any app, hold Fn, speak, and let go.
DaemonFlow lives in the menu bar, with no Dock icon.
- Transcription, cleanup and screen awareness all run on your Mac.
- Audio is kept in memory while you dictate and is never written to disk.
- Vocabulary and history live in
~/Library/Application Support/DaemonFlow/(vocabulary.json,transcripts.json), not encrypted. - Whisper models live in
DaemonFlow/Modelsand cleanup models inDaemonFlow/MLXModelsnext to them. Parakeet is stored by FluidAudio in~/Library/Application Support/FluidAudio/Models. - The only network access is downloading models, from Hugging Face, when you choose one.
- Fn or paste stopped working after an update: DaemonFlow isn't signed with a developer certificate yet, so macOS treats each new build as a different app. Remove DaemonFlow from System Settings → Privacy & Security → Accessibility and Input Monitoring, then add it again.
- "DaemonFlow can't be opened": see step 2 of Getting started.
- First dictation is slow: the models load into memory on first use, and stay loaded after that.
Requires Xcode with the macOS 27 SDK and cmake.
git clone --recurse-submodules https://github.com/kaihere14/DaemonFlow.git
cd DaemonFlow
./scripts/build-whisper.sh # builds whisper.cpp static libs
./scripts/build-and-install.sh # Release build → /Applications/DaemonFlow.app