Voice typing for Windows · runs on your machine · no account
Dictation that gets the hard words right.
Other dictation tools hear sound and guess. Vtypes checks what's already on your screen first, so it knows you said symlinks and not “sink links”.
Real timings from a 2026 desktop CPU, with a hosted model doing the rewrite. Local-only mode trades that speed for never leaving the machine.
A microphone cannot hear spelling.
Every dictation tool turns sound into words. Sound is ambiguous, and a pipeline holding nothing but a waveform resolves the ambiguity by guessing — fluently, which is the dangerous kind. Vtypes reads the visible text of the window you are dictating into and puts it in front of the model that cleans up your transcript. The answer was already on your screen.
- 01
point the CI at the sink links symlinks in the repo
Cursor · CLAUDE.md - 02
restart the cooper netties Kubernetes pods after the deploy
Terminal · kubectl get pods - 03
run pee en pee em pnpm install and try that again
Terminal · pnpm-lock.yaml - 04
the sequel light SQLite migration still needs an index
VS Code · 003_add_index.sql - 05
spin the staging box up on axe your Azure instead
Chrome · portal.azure.com
Screen reading is a switch on the Transcription page, off until you turn it on. What it reads goes to the model endpoint you configured and nowhere else — there is no Vtypes server in this path.
Nothing leaves unless you send it.
Five stages. Three of them physically cannot leave your machine. The other two run locally too, unless you point them at an API — and then they go to your endpoint, on your key. There is no Vtypes account, no Vtypes server, and nothing in the middle to trust.
-
Capture
Hold the shortcut. Audio is buffered in memory.
- optional → your API
Transcribe 312ms
whisper.cpp on your CPU, or an endpoint you chose.
-
Read the screen 8ms
The focused window's visible text, via Windows UI Automation.
- optional → your API
Rewrite 244ms
A local model, or the API key you supplied. Never ours.
-
Paste
Into whatever had focus. Nothing is stored unless you keep history.
Local
Whisper and the rewrite both run on the machine. Works on a plane. Costs nothing per word. Slower on old hardware, and that is the whole trade.
Your own key
Point it at Groq, OpenAI, or anything speaking the same protocol. Faster and sharper. Billed by them, to you, at whatever they charge — we never see the request.
There is nothing to subscribe to.
Vtypes does not sell transcription. It is a client for whatever you point it at — including a model on your own machine, which is the default and costs nothing. If you would rather use a hosted model, you bring the key and you are their customer.
- The app
- free
- No account, no sign-in, no trial clock.
- Local mode
- $0
- Whisper and the rewrite run on your CPU. Nothing is called, so nothing is billed.
- Your own API key
- their rate
- Billed by your provider, to you. We never see the request or the invoice.
A worked example
An hour of continuous dictation is roughly nine thousand words. Hosted speech-to-text is billed by audio duration and the rewrite by token, and dictation is small on both counts — an hour of talking is a few tens of thousands of tokens, most of it the screen context. At the rates the cheap providers published in August 2026 that lands in single-digit cents.
Rates change and this page is not a quote. The reason to trust the shape rather than the number is that we are not in the transaction: check your provider's price list, and that is what you pay.
Say the sentence. Keep the words.
Install it, pick local or your own key, and hold the shortcut. There is no sign-up step because there is no account.
Windows and macOS on Apple Silicon, both at 0.1.28. The macOS build is not signed by Apple, so its first launch needs one extra step. Linux builds from the same codebase and we have not validated it.