Utter
Voice dictation for macOS that runs on your Mac. Hold a key, speak, and the text appears in whatever field you're typing in, cleaned up by a local AI model. No word limits, and nothing leaves your Mac unless you choose the optional cloud cleanup with your own OpenRouter key.
Apple Silicon only.
Install
- Download
Utter-<version>-arm64.dmg from Releases, open it and drag Utter into Applications.
- Open Utter. When macOS says it could not verify the app, click Done, then go to System Settings → Privacy & Security and click Open Anyway. The app is signed ad hoc, not notarized, and since macOS 15 right-click → Open no longer gets past this.
- It lives in the menu bar (🎙), not the Dock. Allow Accessibility when macOS asks (System Settings → Privacy & Security → Accessibility → Utter). The dictation key starts working as soon as you allow it.
- Allow the Microphone the first time you dictate.
Nothing else to install: the speech engines (Phonon and WhisperKit) and the cleanup engine (mlx-lm) ship inside the app. The first run downloads the models, with progress shown in Settings: about 160 MB for speech (646 MB if you dictate in a language Phonon doesn't cover, such as Hindi), and 1.7 GB (English) or 4.3 GB (other languages) for cleanup.
After rebuilding and reinstalling, macOS may ask for Accessibility again.
Build
Building needs Homebrew's whisperkit-cli (copied into the app) and uv (fetches the bundled Python, with mlx-lm and Phonon's fermion-research, minus the PyTorch it only needs on other platforms):
brew install whisperkit-cli
npm install && npm run make
npm run make runs scripts/build-engines.sh and writes out/make/dmg/arm64/Utter-<version>-arm64.dmg.
Use
- Hold the dictation key (Right ⌥ Option by default; change it in the app) and speak. Release to insert the text.
- Double-tap for hands-free; tap again to stop.
- Triple-tap to undo the last dictation.
- ⌃⌘V pastes your last dictation and ⌃⌘C copies it. If nothing that takes text had focus when you finished speaking, the text waits for ⌃⌘V instead of being lost. Change these shortcuts, or add more, under Shortcuts.
- A soft tone plays when dictation starts and stops (Settings → Play sounds). Settings can also open Utter when you log in.
Settings, history and your dictionary are in the app window: click the menu-bar icon → Open Utter. Data is kept in ~/.dictateai (readable only by you).
Speech models
Settings → Speech model → Automatic picks by your dictation language:
| Language | Model | Download |
|---|
| English | Phonon-2 | 158 MB |
| Auto-detect, and 24 other European languages (French, German, Spanish, Russian, Ukrainian…) | Phonon-2.5 | 156 MB |
| Everything else (Hindi, Arabic, Japanese, Chinese…) | Whisper Large v3 Turbo | 646 MB |
Phonon transcribes 7.6 seconds of speech in about 0.15 s on an M-series Mac, and the words in your dictionary are passed to it as hotwords, so it favours your spellings of names and terms. Auto-detect only hears Phonon-2.5's languages; to dictate in Hindi, choose Hindi. Settings can also run Whisper (Small or Large v3 Turbo) for every language.
Phonon-2 and Phonon-2.5 are by Fermion Research, derived from NVIDIA's Parakeet TDT 0.6B v3, and released under CC-BY-4.0. They run through the fermion-research package (Apache 2.0).
Develop
npm run build:engines && npm start
Set UTTER_DEBUG=1 to log dictated text; by default logs show only its length.
npm test runs the unit tests (Node 23.6 or later, which runs TypeScript directly). The start and stop sounds are generated by npm run build:sounds.
Website
The landing page, utter.blinkwiser.com, is site/: one static page, plus two small Vercel functions that look up the newest release on GitHub. /download sends visitors to its .dmg, and the page's buttons show its version and size. To host it on Vercel, import the repository and set Root Directory to site (framework preset: Other, no build command). Elsewhere, the page still works and its buttons open the Releases page. An optional GITHUB_TOKEN environment variable raises GitHub's limit of 60 lookups an hour, which the 10-minute cache rarely reaches.
History
Utter was called DictateAI until October 2026. Its data folder (~/.dictateai) and bundle ID keep that name, so settings, history, models and macOS permissions carry over.
How this version was built, and why: docs/sessions/2026-10-18-session.md.
