How to dictate on Windows without sending your voice to the cloud

If you press Win+H on a laptop with no connection, nothing happens. That is not a bug - it is the clearest possible statement of where your voice goes when you use Windows’ own voice typing. For some work that is fine. For interview recordings, clinical notes, legal drafts or anything under an NDA, it is the whole question.

why Win+H needs the internet

Voice typing sends your audio to Microsoft’s online speech service and gets text back. The recognition happens on their servers, not your CPU, which is why it stops working offline and why it is as good as it is - it is running a far larger model than your laptop would.

What happens to the audio afterwards depends on your privacy settings, your account type and a policy that can be revised. You can go and read those settings. The point is that you would have to keep going back to read them, and the alternative is not having to.

the offline option Windows already has

Voice Access, in Windows 11 22H2 and later, is a different feature from voice typing and runs on-device after a one-time language download. It is built for controlling the whole PC by voice and it dictates as part of that.

Two things to know before you settle on it: it supports a much shorter list of languages than voice typing does, and it is a full voice-control mode rather than a press-to-talk key, which is a different way of working than most people want for dictating a paragraph into an email. If English is your language and you like the modality, it is free, built in, and it never sends audio anywhere. Start it from Accessibility settings.

the third route: a local recognition model

Two open speech models are good enough to run this way on an ordinary laptop, and between them they cover most needs:

• Whisper (OpenAI, open weights). Many languages, gives you word timestamps, can translate into English, and accepts a prompt so you can feed it names and jargon it would otherwise mangle. Slower.
• Parakeet (NVIDIA). Much faster on CPU - well under a second for a typical dictation take - but no timestamps, no translation and no prompt. Fewer languages.

You can run either from a command line yourself; whisper.cpp is a single binary and a model file. What you will not get that way is the thing that makes dictation useful: a global hotkey that types the result into whatever window you are actually in. For that you want an app that wraps the engine. There are several for Windows now, open source and commercial, subscription and one-time.

the four options side by side

how to check a local app is telling the truth

“Private” and “on-device” are marketing words until you test them. Three ways, easiest first:

1. Turn off wi-fi and dictate. A local recognizer keeps working. This catches most of it, and it takes ten seconds. Do it after the model has downloaded - every local app has to fetch its model once.
2. Watch the network while you speak. Open Resource Monitor (resmon), go to the Network tab, find the app, and dictate a long paragraph. Audio going out is not subtle: you are looking for sustained kilobytes, not a few hundred bytes of update check.
3. Block it outbound. Windows Defender Firewall lets you add an outbound rule for one program. If dictation still works with the app blocked, the audio is not going anywhere.

If the app is open source you can go further and read what it sends - but for most people the firewall rule is a stronger guarantee than a privacy policy, because it does not depend on anyone keeping a promise.

what you give up, honestly

A local model is smaller than what a data centre runs. In practice that shows up in three places: proper nouns and product names, heavy accents, and very technical vocabulary. The first is fixable - most local apps let you supply a list of terms and spellings, which works well. The other two are real, and if dictation accuracy is the difference between the tool being used and not, test it on your own voice before deciding.

The bigger practical win is nothing to do with privacy: local dictation does not care about your connection. It works on a train, on hotel wi-fi, and in the room with bad signal where the cloud one silently drops the first half of your sentence.

Who wrote this. We build owntools, which is one of the apps in that last category. The comparison above is written to be useful even if you pick something else - Voice Access in particular is free, already installed, and genuinely on-device.