Native macOS on-device transcription — hold a hotkey, speak, release, text at the cursor.
Nothing is uploaded. Transcription runs locally on Parakeet TDT v3 (via FluidAudio), with WhisperKit as the alternative engine.
Download utt-<version>.dmg from the latest release,
open it, and drag utt.app onto the Applications shortcut inside.
The disk image is signed and notarized, so there is no quarantine flag to clear. docs/INSTALL.md has the three permissions utt needs and how to uninstall.
After that utt updates itself: it looks once a day, shows you what changed, and installs only when you say so.
just install # generate the Xcode project
just run # build and launch
just check # the full gate before committingThe first launch downloads the model (~650 MB) and asks for three permissions: Microphone, Input Monitoring and Accessibility. macOS prompts for each of these exactly once per app, ever — after that the only route is System Settings, which is why utt deep-links there instead of re-prompting.
| Hold the hotkey | record while held, transcribe on release |
| Double-tap it | lock recording on; tap again to stop |
| Escape | cancel without transcribing |
| ⌥⇧V | paste the last transcript again |
The window has two shapes: a floating pill that stays out of the way, and a panel with history and settings. The chevron switches between them.
Recording starts ~0.45 s before the keypress — the engine keeps a rolling buffer, so a word already in flight when you reach for the key still lands in the transcript. That costs a permanently lit microphone indicator while utt is running; the engine stands down on sleep and screen lock, and the whole feature has an off switch in Settings → Recording.
Utt/ SwiftUI app — App, Clients, Design, Features, Views
UttCore/ SwiftPM package: pure logic and all the tests
docs/ hotkey semantics spec, phase 0 spike results
project.yml XcodeGen source of truth (Utt.xcodeproj is generated)
See CLAUDE.md for the architecture and the invariants that are load-bearing.
- Parakeet TDT v3 — NVIDIA NeMo, CC-BY-4.0
- Whisper — OpenAI, MIT, via WhisperKit