Whisper dictation for Windows
Run Whisper on your own PC and type with your voice.
Whisper is OpenAI's open speech recognition model. VinPUT runs it on your PC: you download a model once, choose it under Models, and tap a key to dictate. What you say is typed into the field that has focus.
The sizes VinPUT lists
| Model | Download | Notes |
|---|---|---|
| Whisper large-v3 turbo | 1.6 GB | 99 languages, best accuracy |
| Whisper large-v3 turbo (compact) | 574 MB | Close to full accuracy |
| Whisper base | 148 MB | Fast on any processor, lower accuracy |
Sizes are the model files in millions of bytes, so Windows shows them a few percent lower. To use another Whisper model, choose Add a model at the end of the list. It takes a ggml file (.bin) from Hugging Face or from your PC, and VinPUT checks that it is a Whisper model. Other formats need converting first.
How it runs
Whisper runs in an engine called whisper.cpp, a separate program that VinPUT starts only while a Whisper model is selected. It stays loaded between dictations, sleeps half a minute after the last one and ends after ten minutes without one. The first dictation after a long pause loads the model again.
Whisper types at each pause of 0.6 seconds or more, so your words appear while you speak. On a processor too slow to keep up, the whole dictation is typed when you stop.
If the PC has a graphics card, Settings, Models shows Run on the graphics card, ticked by default. VinPUT then fetches an 18 MB Vulkan build of the engine, which runs on NVIDIA, AMD and Intel cards. With several cards it uses the one with the most memory.
Languages and made-up words
Under Settings, General, Languages, enter codes such as en-US. With one language set, Whisper uses it. With several, Whisper works out which one you are speaking and keeps to your list.
Whisper is known to invent words in silence. VinPUT never types sound names such as [Music] or (applause), and since 0.3.3 local models no longer type "Thank you." after you stop speaking.
If it feels slow
Choose a smaller model such as Whisper base, or switch on the graphics card. The Models page says for each model whether it runs well, may be slow or is too heavy for your PC. More in the documentation.
Questions
Which Whisper size should I pick?
Whisper large-v3 turbo for the best accuracy and 99 languages. The compact version if you want close to that from a smaller download. Whisper base on a slower PC.
Does it type while I speak?
Yes, at each pause of 0.6 seconds or more. On a processor too slow to keep up, it types the whole dictation when you stop.
Do I need a graphics card?
No. Whisper runs on the processor. If the PC has a card, VinPUT uses it by default, and it can be NVIDIA, AMD or Intel.
Can I use my own Whisper model?
Yes, if it is a ggml file ending in .bin. Add it at the end of the Models list.
Related pages
- Parakeet on Windows NVIDIA Parakeet TDT v3 on a processor or any graphics card.
- Offline speech to text What works with no connection, and what needs one.
- Superwhisper for Windows, compared Two dictation apps with local models, side by side.
- Vibe coding with your voice Speak prompts into Cursor, Codex or a terminal.