Lipwise

Speak naturally. Lipwise types what you meant.
Private voice dictation for every app on your computer, with an AI editor that understands “scratch that”.

Download · What it does · Features · Screenshots · FAQ · Build from source

Hold a shortcut, talk, let go. Lipwise transcribes your voice on your own computer and types the result wherever your cursor is: email, Slack, your editor, a browser form. Before it types, an AI editor cleans up what you said and carries out the instructions you gave along the way. Fillers disappear, corrections stick, and “new paragraph” starts a new paragraph.

What it does

Three shortcuts cover everything:

Hold to talk, or tap once to go hands-free until you tap again. Esc cancels. All shortcuts can be changed.

Features

• 🔒 Your voice stays on your computer. Speech recognition runs locally. Only the transcript goes to the AI, and with local AI not even that.
• 🧠 An AI editor in the loop. Understands “scratch that”, “no wait…”, “replace X with Y”, “make it friendlier”, spoken punctuation, lists and translation, plus commands you define yourself.
• 💻 One-click local AI. Lipwise checks your hardware, downloads llama.cpp and a Qwen3.5 model that fits, and runs it in the background. No account, no API key, nothing leaves the machine.
• ☁️ Or bring any model. Claude, OpenAI or any OpenAI-compatible API, Ollama, LM Studio, vLLM or your own llama.cpp server.
• 🎙️ 69 speech models, 100+ languages. Parakeet, Whisper, Canary, Moonshine, SenseVoice, Qwen3-ASR, Voxtral, GigaAM, Granite and more, all through transcribe.cpp. Downloads resume if interrupted and are SHA-256 verified.
• 🗣️ Your words, your commands. Teach it names and jargon, set a writing style (“British spelling, no em dashes”), and add voice commands such as “sign off” → your email closing.
• 🧪 Playground. Type what you'd say and see exactly what would be typed, without touching the microphone.
• 🛟 Never loses a word. If the AI is unreachable, the raw transcript is typed instead, and History shows what happened.
• ✨ Stays out of the way. Lives in the tray, shows a small status pill that never steals focus, restores your clipboard after pasting, and updates itself.
• 🪶 Small and native. Built with Rust and Tauri: about a 5 MB installer, not a 100 MB browser bundle.

Screenshots

Download

Get the latest version from the Releases page.

Then:

1. Open Lipwise and download a speech model. The recommended ones are 200–800 MB.
2. Choose an AI editor: On This Computer for fully private editing, or connect Claude, OpenAI or a local server. You can also turn AI off and use Lipwise as plain dictation.
3. Hold Ctrl+Shift+Space (⌥⇧Space on a Mac) in any app and start talking.

Private by design

• Audio never leaves your computer. Transcription always runs locally.
• The AI sees only the transcript, and only when you've chosen a cloud provider. With local AI, your own server, or AI turned off, nothing is sent anywhere.
• No account, no telemetry. History and settings are stored in your user profile. API keys are kept in the system credential store (Keychain on macOS, Credential Manager on Windows, the Secret Service keyring on Linux), or come from ANTHROPIC_API_KEY / OPENAI_API_KEY.
• Network access is limited to downloading models and llama.cpp, the AI provider you configure, and checking GitHub for updates (which you can turn off).

Local AI

Choose AI Editor → On This Computer → Set Up Local AI, and Lipwise does the rest: it detects your GPU and memory, downloads a matching llama.cpp build and model (checksums verified), and runs it hidden in the background.

It uses Metal on Apple Silicon, Vulkan on any GPU with 2 GB or more of its own memory (NVIDIA, AMD or Intel), and the CPU otherwise. The model starts with Lipwise, stops when you quit, and Remove Local AI deletes everything it downloaded.

How it works

 shortcut ──► microphone ──► speech model ──► AI editor ──► typed into your app
                             (on device)          │
                                                  └─ AI off or unreachable? The raw transcript is typed instead.

FAQ

How is this different from Handy?

Handy is excellent local dictation, and Lipwise builds on its model catalog. Lipwise adds the AI editing step: spoken corrections, formatting, rewrites, command mode on selected text, and custom voice commands. Plain dictation still works exactly like Handy, and the default shortcuts differ so both apps can run side by side.

Does it work offline?

Yes. Once a speech model is downloaded, transcription is entirely offline. With the local AI editor (or AI turned off), the whole pipeline runs without an internet connection.

Which AI should I pick?

For privacy, use On This Computer. The 4B model handles every built-in command and is quick on a GPU or Apple Silicon. For the best results on long rewrites and translations, use Claude or the 9B local model. Any model that follows instructions well will work.

What if the AI misunderstands me?

Use Plain dictation for anything that must be verbatim. To stop ordinary sentences being read as commands, set a wake word in Voice Commands; only speech that starts with it is treated as an instruction. Every dictation in History shows what was heard next to what was typed.

How do updates work?

Lipwise checks GitHub Releases now and then, downloads new versions in the background, and installs them the next time it restarts. You can turn this off in Settings → Updates.

Contributing

Bug reports, model results on your hardware, and pull requests are very welcome. See CONTRIBUTING.md for building from source, running the tests and the project layout.

Credits

The speech model catalog and the transcribe.cpp integration approach come from Handy by CJ Pais (MIT). Local AI runs on llama.cpp and the Qwen3.5 models.

License

MIT