Press a shortcut in any app, speak, and clean text lands where you’re typing. Free, open source, runs on your computer.
Apple Silicon recommended · No account · No telemetry · or build from source
um yes let's meet at 3 no actually 4
Yes! Let's meet at 4.
Press Fn and talk. The pill listens, and a quick check makes sure it heard a voice, not a key click.
Speech recognition turns your audio into words, on your computer or on a server you choose.
A small language model drops the fillers, applies your corrections and adds punctuation.
The clean text is pasted where your cursor is, and your clipboard is put back as it was.
Most people type around 50 words a minute and speak around 150. Typelite turns that speed into finished text, so what you say is ready to send, without the ums or the retyping.
Typelite shows your own speaking and typing speeds in Insights, kept on your computer.
Rates are rounded down and vary from person to person. Typing: an average of 52 words per minute across 168,000 volunteers (Dhakal et al., CHI 2018). Speaking: about 196 words per minute measured in telephone conversations (Yuan, Liberman and Cieri, Interspeech 2006) and 153 for English speech input on a phone (Ruan et al., 2016); we use a conservative 150.
Dictate, translate or ask, from whatever you’re typing in. A small pill at the bottom of your screen shows what’s happening and never steals focus.
Talk the way you talk. Typelite writes it the way you would have typed it, and pastes it where your cursor is.
Speak in your own language and the text arrives in the one you need. The recording keeps going while you switch.
Highlight a paragraph and ask for a summary, a rewrite or an explanation, without leaving the app you’re in.
Thinking
Cantonese speakers mix English into almost every sentence; British English spells things its own way. Each of your languages can carry its own instructions, a language preset, and a small prompt router adds only the one for the language you just spoke.
我哋聽日開會之前，你可唔可以check下個deck？
Two Cantonese hint words (聽日, 唔): the Cantonese preset is added.
Presets are small Markdown files in the repository, written and reviewed by people who speak the language. Preview one, use it, edit it: your edits are kept, and updates are offered, never forced.
Colloquial written Cantonese as Hong Kong people type it, with Hong Kong code-mixing.
Clear, natural English that keeps the speaker's tone, with spelling and words per region.
Natural Taiwan Mandarin in Traditional characters, with Taiwan vocabulary and punctuation.
If you speak a language well, you can make Typelite write it better.
Contribute a presetBrowse the preset catalogue or read how languages work.
Typelite works in two steps, speech recognition and AI polish, and each runs where you choose. Mix them freely.
No servers to install. On Apple Silicon, whisper.cpp runs inside the app and llama.cpp’s server ships with it. Models are picked for your hardware and checked against a SHA-256 before use.
Prefer a GPU computer on your network, or a cloud key? Typelite talks to any OpenAI-compatible speech or chat server, and to the Qwen Cloud and ElevenLabs speech APIs.
Cloud services are used only when you add one with your own key; they then receive your audio or text. Nothing leaves your machines otherwise.
Read the setup guidesTypeless is a polished commercial product and a great source of ideas. Typelite is for people who want to own the whole pipeline and tune it for their languages.
| Aspect | Typeless | Typelite |
|---|---|---|
| Source | Closed | Open source (MIT) |
| Price | Free tier and paid plans | Free |
| Account | Sign-in required | None |
| Where speech and AI run | The Typeless cloud | On your computer, your own server, or a cloud service you pick |
| Languages | Built in | Per-language presets from a public library, plus a router |
| Translate targets | Up to 3, switch from the pill | Up to 3, switch with a key or the pill while recording |
Based on public information about Typeless; it may change. Typelite is an independent project and not affiliated with Typeless.
Every line is on GitHub: read it, build it, change it. Code, service cards, language presets and docs are all welcome.
There is no signed release yet. You’ll need Rust, Node.js 20+, the Xcode Command Line Tools and CMake; details in CONTRIBUTING.md.
Typelite$ git clone https://github.com/sennett-lau/typelite.git
$ cd typelite
$ npm ci
$ npm run build:app # builds llama-server, then the app bundle
$ open src-tauri/target/release/bundle/macos/Typelite.app
Tauri 2 with React and Rust. Fix a bug or build a feature; larger ones start with a short plan.
ContributeDocument a speech or AI service that already works, so others can connect it in a minute.
ContributeSpeak a language well? Teach Typelite how to write it, in a small Markdown file.
ContributeGuides for setups, models and troubleshooting. Clearer docs help everyone.
ContributeNothing to sign up for, nothing to sign in to. Install it and speak.
No analytics, no tracking. Logs keep timings and errors, never your text.
Audio, transcripts and answers are never stored. Insights keep timings only, and one click clears them.
The built-in models, your own server, or a service you picked with your own key. Nowhere else.
Free forever, open source, and it runs on your own computer.