LipwiseLipwise types what you meant: private voice dictation with an AI editor, for every app on your computer
Voice dictation for every app on your computer. Hold a shortcut, talk, let go: your speech is transcribed on your own machine, and an editor turns what you said into what you meant.
§ 1
The proof
What was said, marked up the way Lipwise’s editor reads it, then typed.
Heard
Typed
- um struck: filler or a false start
- question mark said, then set as punctuation
- make it polite an instruction, carried out
- scratch that stet: sounds like one, but it’s your content
§ 2
Three shortcuts
Hold one in any app, talk, let go. Or tap it once to go hands-free until you tap again; Esc cancels. Played out below with a scripted voice.
- Dictate talk, with instructions mixed in
- Command select text, say what to do with it
- Plain exactly what you said, no AI
Messageany app
Idle
Said:
§ 3
Where your words go
Audio never leaves your computer. Where the text goes depends on the editor you pick; here are all three.
Your computer
- Microphone16 kHz, in memory
- Speech modelone of 69, runs locally
- AI editorQwen3.5 via llama.cppsends the transcript outskipped
- Your apppasted, clipboard restored
The internet
AI providerClaude, OpenAI, or any OpenAI-compatible APInot used
Leaves your computer: nothing.
- No account and no telemetry. History and settings live in your user profile.
- API keys are kept in Keychain, Windows Credential Manager or the Secret Service keyring.
- Network use is limited to downloading models, the provider you set, and update checks you can switch off.
- Local AI is one click: Lipwise checks your GPU and memory, fetches a matching llama.cpp and a Qwen3.5 model (4B, 2.7 GB or 9B, 5.7 GB), verifies them and runs them in the background.
§ 4
A phrasebook
Things you can say while dictating. No syntax to learn; say them the way you would to a person taking notes.
- scratch that phr.
- Takes back what you just said. “the price is 40 dollars scratch that 45 dollars” The price is 45 dollars. See also: delete that, never mind, forget all of it.
- no wait interj.
- Swaps the last thing for what follows. “send it to john actually send it to maria” Send it to Maria.
- new paragraph n.
- Starts one. Likewise new line, bullet point, numbered list, make it a list. “agenda colon numbered list budget hiring roadmap” Agenda:
1. Budget
2. Hiring
3. Roadmap - question mark n.
- Becomes ? And so on for comma, period, colon, open parenthesis, put that in quotes, all caps.
- make it … v.
- Said at the end, rewrites everything before it: formal, friendlier, shorter, more casual, a short tweet, a proper email. “thanks for the update I'll look at it tomorrow rewrite that more formally” Thank you for the update. I will review it tomorrow.
- translate this into … v.
- Any language your model writes. “I'll be a few minutes late, sorry. Translate that into German.” Tut mir leid, ich komme ein paar Minuten später.
- replace … with … v.
- Finds and swaps, without you reaching for the mouse.
- sign off n., yours
- A command you define. Teach Lipwise your own phrases, names and jargon, and a house style such as British spelling, no em dashes. “hi team the release is on track sign off” Hi team, the release is on track.
Best regards,
Alex - wake word n., optional
- Set one, and only speech that starts with it counts as an instruction. Everything else is typed as content, however much it sounds like a command.
§ 5
Contact sheet
The settings window, which you’ll rarely open. Mostly Lipwise lives in the tray.
§ 6
Specifications
- Installer
- About 5 MB. Native Rust and Tauri, not a browser bundle.
- Speech models
- 69 through transcribe.cpp: Parakeet, Whisper, Canary, Moonshine, SenseVoice, Qwen3-ASR, Voxtral, GigaAM, Granite and more.
- Languages
- 100+
- Local AI
- Qwen3.5 4B or 9B on llama.cpp. Metal on Apple Silicon, Vulkan on GPUs with 2 GB or more, CPU otherwise.
- Cloud AI
- Claude, OpenAI, Ollama, LM Studio, vLLM, a llama.cpp server, or any OpenAI-compatible endpoint.
- Fallback
- If the AI is unreachable, the raw transcript is typed. No words lost.
- Playground
- Type what you’d say and see what would be typed, microphone untouched.
- Account
- None
- Trackers & telemetry
- None
- Price
- Free. MIT licensed.
§ 7
Download
From GitHub Releases. Lipwise keeps itself up to date after that.
| Platform | File | Size | Link |
|---|---|---|---|
| Windows 10, 11 | x64 installer (.exe) | — | Download |
| macOS 11+, Apple Silicon | disk image (.dmg) | — | Download |
| Linux, Debian/Ubuntu | package (.deb) | — | Download |
| Linux, Fedora/openSUSE | package (.rpm) | — | Download |
| Linux, any distribution | AppImage | — | Download |
- Pick a speech model in the app. The recommended ones are 200–800 MB.
- Pick an editor: on this computer, a provider, or off for plain dictation.
- Hold Ctrl Shift Space in any app and talk.
Windows: if SmartScreen warns about a new app, choose More info → Run anyway. macOS: allow Microphone and Accessibility on first use; Accessibility is what lets Lipwise type into other apps. Linux: X11 works best; on Wayland, shortcuts and typing depend on the compositor.
§ 8
Questions
Does it work offline?
Yes. Once a speech model is downloaded, transcription is fully offline, and with the local editor (or none) so is everything else.
What if it mistakes my words for an instruction?
Use Plain for anything that has to be verbatim, or set a wake word so only speech that starts with it is treated as an instruction. History shows what was heard beside what was typed, every time.
Which editor should I choose?
For privacy, On This Computer: the 4B model handles every built-in instruction and is quick on a GPU or Apple Silicon. For the best long rewrites and translations, Claude or the 9B local model. Any model that follows instructions well will do.
How is this different from Handy?
Handy is excellent local dictation, and Lipwise builds on its model catalog. Lipwise adds the editing step: corrections, formatting, rewrites, Command mode on selected text, and your own voice commands. The default shortcuts differ, so both can run side by side.
How does it update?
It checks GitHub Releases now and then, downloads quietly, and installs on the next restart. You can turn this off in Settings → Updates.
Can I build it myself?
Yes: Tauri 2, a Rust core and a React interface. CONTRIBUTING.md covers building, tests and the layout. Results on your hardware and pull requests are welcome.