Документация
How to dictate offline on Windows 11
There are four ways to turn speech into text on a Windows 11 machine without a connection, and only one of them is what most people try first. This page walks through all four, says what each sends to a server, and ends with a setup that works on a plane.
Обновлено · Vitalii Yemets
Why offline matters
Two reasons, and they are different. The first is the connection: a train, a plane, a site with no signal, a network that blocks everything but mail. The second is privacy: what you dictate may be a patient's name, a contract, a password you are about to paste. When recognition runs on the computer, neither reason applies. When it runs in a cloud, both do, however well the cloud is certified.
Option 1: Win+H voice typing
Press Win+H in any text field and talk. It is free, already installed on every Windows 10 and 11 machine, and covers 43 languages with optional automatic punctuation. It is not offline: Microsoft's own page says voice typing needs an internet connection and uses online speech recognition on Azure. There is no dictionary of your own words. For a sentence a day it is enough; for a page a day, the corrections add up, and every word goes to Microsoft.
Where it fits next to a local program is on the Win+H comparison page.
Option 2: Voice Access
Voice Access is the newer tool in Windows 11: on-device recognition after a language pack download, so it works without a connection, and it goes beyond dictation to controlling the PC by voice: open apps, click, scroll. It covers six languages: English, Spanish, French, German, Chinese and Japanese. If you need control by voice, it is the right tool, and it runs side by side with everything below. If you write in Polish, Ukrainian or Italian, it does not help; and the old Windows Speech Recognition it replaces is deprecated.
Option 3: a local model in a dictation program
Open speech models such as Whisper, Parakeet and Nemotron run on an ordinary processor, and a handful of programs wrap them into push-to-talk dictation for any window. HoldToType is one of them: hold Ctrl+Win, speak, let go, and the text lands at the caret; nine models to choose from, 99 languages with Whisper, a dictionary, translation, live text on a plate, and nothing that ever goes online. Handy and OpenWhispr's local mode do the core of the same thing on Mac and Linux too; the comparison pages say where each one is better.
This is the option for every language, for privacy you can verify in a log, and for a dictionary of your own words. What it costs is a download of a few hundred megabytes and a processor that is not from the previous decade.
Option 4: Whisper on files
If what you have is a recording rather than a live voice, the Whisper command-line tools and desktop wrappers turn audio files into text offline. They do not type into other programs; that is a different job. HoldToType does not transcribe files either: it listens live.
Side by side
- Works without a connection: Win+H no; Voice Access yes; a local program yes; Whisper on files yes.
- Languages offline: Voice Access six; a local program 99 with Whisper, faster narrow models for 25 European languages and Russian.
- Your own dictionary: Win+H no; Voice Access not documented; a local program yes.
- Controls the PC by voice: only Voice Access.
- Where the audio goes: Win+H to Azure; the others nowhere.
A setup that works on a plane
- Install HoldToType from the download section; the wizard downloads a model for your language while you still have a connection.
- Pick a model on the models page: Parakeet for speed, Nemotron 3.5 for live text, Whisper Medium if you also want translation.
- Put the caret in a text field, hold Ctrl+Win, say a sentence, let go. Open the log from the System tab: the count of network requests for that dictation is zero.
- If you also want to click by voice, turn on Voice Access; the two do not conflict.
Questions
Can Win+H be made offline?
Microsoft documents voice typing as an online feature. The on-device path Microsoft offers is Voice Access, in six languages.
How much disk does a local model take?
From 74 MB for the smallest Whisper to about a gigabyte for the most accurate one; the usual pick for a European language is around 650 MB.
Is a local model as accurate as the cloud?
On everyday speech in a major language, close. The cloud wins on rare accents and on rewriting; a local program wins on names you put in its dictionary and on privacy.
Dictate offline today
A minute to install and a model download, then no connection needed again.