NabuVoice · Dictation that stays put
Hold a key. Talk.It types for you.
Hold your hotkey anywhere on the desktop, speak, and release. The words land at your cursor in whatever app already had focus — transcribed by a model running on your own hardware.
MACOS 12+ · WINDOWS IN PROGRESSMIT LICENSED
01What it does
Everything you need to stop typing.
A fast, private dictation workflow that meets you in the apps you already have open.
Push-to-talk, anywhere
Hold your hotkey — fn by default — speak, and release. A native helper owns the key tap at the OS level, so it works without switching windows or clicking into anything.
Works in every app
Text is inserted at your cursor through the accessibility layer, with a clipboard paste as the fallback. Mail, chat, browser, editor — if you can type there, you can dictate there.
Speech models on your machine
Pick a model in the app, download it once, and transcription happens on-device from then on. No account to create and no server in the path.
Cleanup by a local LLM
Grammar and formatting are tidied by a small language model running through llama.cpp on Metal — on the same machine, straight after the transcript lands.
Vocabulary and snippets
Teach it the names, jargon and acronyms you actually use so they come out spelled right, and expand a spoken shortcut into a full block of text.
Notes and per-app rules
Capture a thought as a note inside the app when you do not want to dictate into something else, and set personalization rules that change behaviour per application.
Five interface languages
The app itself ships in English, German, Spanish, Japanese and Traditional Chinese, and the Whisper models understand far more spoken languages than that.
02Privacy
Private by design. Local by default.
A dictation app hears everything you say, including the half-sentences you throw away. That is exactly why this one keeps the audio on the machine that heard it. There is no account to create, no server to trust, and nothing to switch off before you say something private.
- Your voice
- Models on your device
- Text in your app
The whole pipeline runs with the network off
Of your audio stays on-device when you transcribe with a local model.
Local speech models, from Parakeet TDT v3 to seven Whisper variants.
Local correction models running through llama.cpp on Metal.
03The models
Eighteen models. All of them yours.
Nothing here is a hosted endpoint with a friendly name. These are weights that land on your disk, get checked, and then run without asking anyone's permission.
Speech models
The transcriber. Pick one in the app, download it once, and it runs on your hardware from then on.
- Parakeet TDT v3Default
- Parakeet TDT v2
- SenseVoice
- Moonshine Base
- Canary 180M Flash
- Canary 1B v2
- Seven Whisper variants
Correction models
The proofreader. A small language model tidies grammar and formatting through llama.cpp on Metal, listed here smallest first.
- Qwen2.5 1.5BDefault
- SmolLM2 1.7B
- Qwen3 4B
- Qwen3 8B
- Qwen3 14B
Every model artifact carries a sha256 checksum and is verified after download, so a corrupted or substituted file never gets loaded.
04Open source
Built in the open.
MIT, not source-available
The permissive one. Use it, fork it, ship something on top of it. NabuCode — the other app in the suite — is proprietary; this one is not.
Free, with nothing behind a gate
No accounts, no subscription, no pro tier holding the useful half hostage. Everything the app does, it does for everybody.
Auditable rather than promised
The privacy story on this page is a claim about code. When the repository is public it becomes something you can check instead of something you have to believe.
Source
The repository is not public yet, so there is nothing here to link to. It goes up with the first release, under the licence above.
05FAQ
Asked and answered.
Early access
One email, when there is a build to install.