holdtotypeInstall and first run
Everything needed to get from a download to a working shortcut. Nothing here requires an account, and none of it needs a network connection after the model is on disk.
Download and install
Two builds, same program. Pick whichever suits the machine.
installer
Puts the program in a folder of your choosing, makes a shortcut and an uninstall entry. It can also download a recognition model during the installation, which saves doing it on first run.
holdtotype-setup.exe silent install, with a folder and a model: holdtotype-setup.exe -silent -dir "C:\Apps\HoldToType" -model small uninstall without questions: holdtotype-setup.exe -uninstall -silent
portable
Unpack holdtotype-portable.zip anywhere, including a USB stick, and run holdtotype.exe. Nothing is written outside that folder and no administrator rights are needed.
First launch
The first run opens a wizard of five short steps. Each can be skipped, and none of them asks for anything beyond what the program needs to work.
- Interface language — eight are available, or it follows Windows.
- Dictation language — the model for it is chosen and downloaded here, with a progress bar.
- Shortcut and microphone — a live level bar shows the microphone is actually heard.
- A field to try it in — a real dictation, into the wizard itself.
- Start with Windows — off unless you say otherwise.
Windows will ask once for microphone permission. If it was refused, the Microphone section says so and links to the system setting.
Choosing a model
The wizard picks a sensible one, and it can be changed at any time in Languages & models. What the choice actually decides:
| If you want | Pick | Why |
|---|---|---|
| a safe default for any language | Whisper Medium q5 | Covers everything and translates into English by itself. About 1.5 GB. |
| the lowest latency | a narrow model for your language | GigaAM v3 turned an 11-second phrase into text in 0.47 s against Whisper's 11.6 s, in a third of the memory. |
| words while you speak | Nemotron 3.5 streaming | Partial text appears on the plate a few times a second, 40 languages. |
| a weak machine | Whisper Base or Moonshine | Small and quick; less accurate on long sentences. |
The catalogue shows the memory each model will actually take, measured against what is free right now, and says plainly whether it can translate.
Your first dictation
- Put the caret where text can be typed — an editor, a browser field, a terminal.
- Hold ctrl + win. A plate appears with a recording dot, a clock and a level bar.
- Say a sentence. Speak normally; there is no need to dictate punctuation if the model adds it.
- Let go. The plate shows recognition, then insertion.
- The text lands at the caret, and your clipboard is restored to what it held before.
0 network requests for that dictation.Shortcuts
| Action | Keys | Notes |
|---|---|---|
| Dictate | ctrl + win | Hold by default; a toggle mode is available in Controls & behaviour. |
| Cancel the current dictation | Esc | While the plate is up. Nothing is inserted. |
| Dictate with a specific prompt | your own | Each post-processing prompt can carry its own combination. |
| Dictate and translate | your own | Set in the Translation block. |
| Find a setting | ctrl + K | Inside the settings window; searches the manual too. |
| Answer the plate | 1…9, Enter | When it asks which language to translate into. |
Any combination can be reassigned. The field warns when Windows has already claimed the one you picked.
Questions people ask
Does it work without an internet connection?
Completely, once a model is on disk. Recognition, translation by the model, prompts and every rule run on your own machine. The only things that need the network are downloading a model and checking for an update, and both happen when you ask.
Do I need a graphics card?
No. Everything runs on the processor. A narrow model turns a normal sentence into text in well under a second on an ordinary laptop; the large Whisper models are slower and want more memory.
Which languages can it dictate?
Whisper covers about a hundred. Beyond it the catalogue has models for particular languages and families: 25 European languages in one, about 30 in another, and single-language models that are faster and more accurate than anything general.
How much disk does it take?
The program with its engines is about 100 MB. Models are what take room: from 150 MB for the smallest to about 1.5 GB for Whisper Medium. You download only the ones you use.
Will my antivirus complain?
It might. The builds are not code-signed yet, so Windows warns about an unknown publisher. Every release publishes the SHA-256 of both files so you can check what you downloaded, and signing is on the roadmap along with the Microsoft Store.
Can I run it from a USB stick?
Yes. The portable build writes nothing outside its own folder — settings, models and history all live there. Copy the folder to another machine and it carries on.
What happens to my voice?
It is held in memory while you hold the keys and dropped as soon as it has become text. It is never written to disk as a file and never uploaded, unless you deliberately set up recognition on another computer.
What does it cost?
Nothing. MIT licence, no account, no subscription, no paid tier waiting behind a feature.
Troubleshooting
the plate does not appear
Another program holds the same shortcut. Change it in Controls & behaviour. If the program was started as administrator and the target window was not — or the other way round — Windows will not deliver the keys; run both at the same level.
the plate appears but the level stays flat
Wrong device or a muted microphone. The Microphone section has a live meter and a record-and-play test.
recognition is unavailable
Either no model is installed for that language, or the recognition server was pointed at another computer whose address is empty. The Status screen names which one and takes you there.
an engine will not start
Since 0.75.4 the program checks the port before starting an engine and refuses if something else holds it. The message names the port; change it in System, or stop the other program. Two copies of HoldToType on one machine need different ports.
read the log
Tray menu → Open log. It is plain text, written in English whatever the interface language is, and it counts characters instead of quoting what you dictated. The last twenty lines usually name the cause.
2026-09-01 11:00:25 ready: hotkey=ctrl+win engine=sherpa model=models/gigaam-v3 2026-09-01 11:02:41 dictation: 11.4 s captured, 0 network requests 2026-09-01 11:02:42 replacements: 62 -> 64 characters 2026-09-01 11:02:42 pasted 64 characters into code.exe