Offline speech to text for Windows
Dictation that keeps working when the network does not.
Once a speech model is on your PC, VinPUT needs no internet connection to dictate. The audio is recognized on your machine and the words are typed into whichever text field has focus, in any Windows program.
What works with no connection
Dictation with any model under On this PC, the hotkeys, the dictionary and snippets, spoken formatting and voice memos. Cleanup works too if you use a text model on your PC (Gemma 4 or Qwen3.5), and then the text never leaves the machine either.
If you use Plus or Pro and the connection drops, VinPUT finishes the dictation with the model you used last on that PC, if one is installed. Nothing is downloaded for this.
What needs the internet once
A model is a download: the model file, plus a small engine that runs it. The engine is 8 MB for the processor, 18 MB for a graphics card, and 33 MB for Qwen3-ASR. Every file is checked against a pinned SHA-256 before it is used, and an interrupted download carries on where it stopped.
VinPUT also checks for a newer version at start and every 6 hours. Switch that off under Settings, About. Plus, Pro and the two Google models need a connection while you dictate, so they are not offline options.
The models
| Model | Download | Notes |
|---|---|---|
| Whisper base | 148 MB | Fast on any processor, lower accuracy. Types at each pause. |
| Whisper large-v3 turbo, compact | 574 MB | Close to full accuracy. Types at each pause. |
| Whisper large-v3 turbo | 1.6 GB | 99 languages, best accuracy. Types at each pause. |
| Parakeet TDT v3 | 669 MB | 25 European languages. Types when you stop. |
| Qwen3-ASR 0.6B | 1.0 GB | 30 languages, not Slovenian. Fast on any processor. |
| Qwen3-ASR 1.7B | 2.5 GB | 30 languages, not Slovenian. Quick without a graphics card. |
Each row on the Models page also says how that model would run on your PC: well, slowly or too heavy.
Checking it yourself
Download a model while you are online, click Use, then disconnect and dictate. If words appear, the recognition happened on your PC. The privacy policy lists the few connections VinPUT makes.
Questions
Does offline speech to text need the internet?
Not while you dictate with a model on your PC. You need it once to download the model, and for the update check if you leave it on.
Which model should I download first?
Look at the Models page, which marks each model as running well, slowly or too heavy for your PC. For most languages, Whisper large-v3 turbo (compact) is a middle choice at 574 MB.
Does my voice leave my PC with a local model?
No. The audio is processed on your PC and sent nowhere.
Does cleanup work offline?
Yes, with a text model on your PC. Cleanup with Plus or Pro sends the text to VinPUT's service in the EU, so it needs the internet.
Related pages
- Whisper dictation The Whisper sizes VinPUT lists and how they run.
- Parakeet on Windows NVIDIA Parakeet TDT v3 on a processor or any graphics card.
- Dragon alternative for Windows What VinPUT covers if you dictate on Windows.
- Voice typing on Windows 11 and Win+H What the built-in option does, and when it is enough.