Press a hotkey, speak, and press it again. Clean, well-formatted text appears in the app you're typing into.
Speech is transcribed on your Mac with whisper.cpp. The cleanup
(filler words, punctuation, numbers, lists, email addresses) runs through the
Claude Code CLI you're already signed in to. That uses your Claude plan's
normal usage limits, with no API key. You can also clean up with any OpenAI-compatible endpoint (Ollama, LM Studio,
OpenAI, Groq…) or fully offline with S1-mini. On macOS 26 and later, Apple's own on-device engines are options too:
SpeechAnalyzer for speech and Apple Intelligence for cleanup.
Early software. This is a young project with one maintainer, tested mainly on one Mac (M3 Pro). It isn't notarized
by Apple yet, so the first launch needs Open Anyway. Bug reports are very welcome.
What it does
- Any app: Slack, Mail, VS Code, the terminal, the browser. The text goes where you started dictating: if you switch
away meanwhile, it's copied instead, and it's never pasted into a password field. Your clipboard comes back afterwards.
- Cleanup that keeps your meaning:
- Removes "um" and false starts, and applies self-corrections ("Friday, no wait, Thursday").
- Writes $15,400 · 15% · 2:30 PM · [email protected] properly, and formats lists and paragraphs.
- If a cleanup ever drops a number or a "not", Whisper's own text is pasted instead.
- ⌃⌥Z swaps the last paste between the cleaned text and Whisper's text.
- Command Mode (⌃⌥⇧Space): select text, press it, and say how to change it: "make this shorter and more polite",
"turn this into bullet points", "translate to Spanish", "it's T-O-N-I". The selection is replaced, and ⌘Z brings it back.
- Follow up with "shorter still" or "go back to the original".
- With nothing selected, it edits what you just dictated, or writes new text at the cursor ("write a two-line thank-you
to the team").
- For text you can't edit (a web page, a PDF, a terminal), the answer goes to the clipboard.
- App-aware modes: casual for chat, paragraphs for email, exact identifiers for code, lists for notes.
- Dictionary for names and terms, a floating indicator (listening, transcribing, polishing, pasted), a microphone
picker, and clear messages when something goes wrong ("Claude limit reached · resets 3:45 PM").
- Nothing lost: if cleanup fails, Apple Intelligence (macOS 26+) or S1-mini cleans up on your Mac instead, and failing
that you get Whisper's text.
Which install is for me?
| You want | Do this | You need |
|---|
| The app (almost everyone) | Download the DMG | Nothing else: no Homebrew, no Terminal |
| To work on the code | Build from source | Xcode Command Line Tools, cmake |
Only the dictate terminal command | ./scripts/install.sh (guide) | brew install sox whisper-cpp |
Install
Requires macOS 13.3 or later on Apple Silicon.
- Download
OpenVoiceType-<version>.dmg from the latest release,
open it, and drag OpenVoiceType to Applications.
- Open it. The first time, macOS says it can't check the app, because it isn't notarized by Apple. Click Done, open
System Settings → Privacy & Security, and click Open Anyway next to "OpenVoiceType was blocked". Once only.
- The setup window walks you through the rest:
- the Microphone and Accessibility permissions (Accessibility is for pasting);
- a speech model (the compressed one, 574 MB, is recommended);
- cleanup: it finds Claude Code, or installs it with Anthropic's installer and signs you in, in Terminal, where you can
see what runs. It also offers S1-mini (484 MB), so dictation works right away, even offline;
- Try it: a box to dictate into.
The models come from Hugging Face (whisper.cpp,
S1-mini), pinned and checked against their SHA-256.
Build from source
git clone https://github.com/mahfuzur/openvoicetype.git && cd openvoicetype
xcode-select --install # if you don't have the Command Line Tools (full Xcode isn't needed)
brew install cmake
./scripts/setup-signing.sh # once: a local signing identity, so macOS keeps permissions across rebuilds
./scripts/build-app.sh --install # builds whisper.cpp and llama.cpp into the app (about 3 minutes the first time)
Apple's on-device engines (app/AppleHelper) need Swift 6.2 and the macOS 26 SDK (Command Line Tools 26 or later). Older
tools build the app without them.
Using it
⌃⌥Space starts and stops (or hold it while you speak), Esc cancels, ⌃⌥⇧Space is Command Mode, and ⌃⌥Z
swaps the last paste. Everything else is in Settings (⌘, from the menu-bar icon). Modes, the dictionary, cleanup engines, the command line and
troubleshooting are in the Guide.
How well it works
Measured on 2026-09-25 on an M3 Pro (18 GB) with evals/run.py, on 25 fixed test cases. Seventeen are
formatting cases (lists, numbers, corrections, email addresses, code). Eight are safety cases: a number or a "not" must
survive, and a dictated instruction ("ignore previous instructions…", "you are now a pirate…") must come back as text, not
be carried out.
| Cleanup engine | Formatting | Safety | Median time per cleanup |
|---|
| Claude Haiku (default) | 17/17 (16/17 in some runs) | 8/8 | 3.4 s with the CLI's startup; about 1.2 s pre-started, as in the app |
| S1-mini (on this Mac, offline) | 8/17 | 7/8 | 0.7 s |
| Apple Intelligence (on this Mac, offline; macOS 26+, measured 2026-10-18 on macOS 27) | 8/17 | 7/8 | 1.8 s |
| Ollama llama3.2, 3B (on this Mac) | 9/17 | 3/8 | 1.8 s |
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
