Can speech to text run on an air-gapped computer?

Yes, but only for a tool that does all of its recognition on the computer and has everything it needs installed before the network is removed. An air-gapped computer has no automated connection to other systems, so any tool that sends audio to a server stops working. What runs depends on what was installed, which languages it covers and whether setup was finished while the machine was still online. Our answer on confidential documents covers the privacy side of the same decision, and this page covers the technical one.

An air gap rules out any tool that calls a server

NIST's glossary defines an air gap as an interface between two systems that are not physically connected and whose logical connection is not automated ¹. Data crosses that boundary only manually, under human control. A dictation tool that needs a server cannot work across the boundary, and it cannot work on the isolated machine either.

That leaves on-device speech recognition. Apple's WWDC 2025 session on its SpeechAnalyzer interface says that transcription is entirely on the device, with one condition: the language models must be downloaded first ². That session describes a developer API. Our sources do not say which engine Mac Dictation uses, so check the behavior of the specific app on the specific Mac.

Built-in tools split between setup that needs the internet and use that does not

Windows shows the split clearly. Microsoft says Voice Access lets you control your PC and write text by voice without an internet connection ³. The first-time setup is different: Microsoft says you need an internet connection to complete the initial setup of Voice Access ⁴. Finish setup on a connected machine, confirm that it works, and only then isolate it. Microsoft's pages do not say whether later language-file updates need a connection, so test that before you depend on it.

Do not confuse Voice Access with voice typing. The Windows voice typing tool, started with Windows key and H, needs an internet connection before it will type ⁵. For a sealed machine, it is not an option. Our answer on whether Windows voice typing works offline explains the difference in more detail.

Each language needs its own offline files

Offline support is set per language, not per product. In Chromium's source for ChromeOS dictation, the code comments say that the languages in the speech list are, by default, not supported offline, and that offline languages are added only when the on-device installer reports them as available ⁶. That source is a snapshot at one commit, but it shows the design: a product can list a language and still have no offline support for it.

Storage matters too. A January 2025 estimate in a Chromium discussion put each on-device speech recognition language pack for Chrome on Windows, Mac and Linux at about 60 MB ⁷. That is small per language, but the sizes add up across several languages, and every file has to pass through whatever transfer process your environment allows.

Developer services for disconnected use need approval

Some developer products can run without a connection, but only after an approval step. Microsoft says that to run its Speech containers in environments disconnected from the internet, you must submit a request form and wait for approval ⁸. The same page marks the containers as in public preview, so the terms can change. For a desktop dictation rollout, this is a developer route rather than an app you install.

It depends on the language, the vocabulary and the updates

• Accuracy. Our sources do not measure how well offline engines handle technical, legal or classified vocabulary. Test your own terms, and proper names, before you rely on the output.
• Languages. Check each language you need, not the product name. A tool that works in English may have no offline pack for your second language.
• Updates. Each engine update and language file has to go through the same transfer process as the first install, or the machine stays on an old version.

What to do

1. Define "air-gapped" for your site. Use the NIST definition in your policy, so that everyone applies the same test ¹.
2. Install and set up while connected. Finish any first-time setup, download the language packs and models, and write down exactly what you installed and where it came from.
3. Move the files with approved media. Check what arrives on the isolated machine against your list before you install it.
4. Test with the network disconnected. Dictate a passage in each language. If the text appears, recognition works offline for that language. If it stalls or fails, the tool is not offline-capable for that language.
5. Reject any tool that needs a server. If setup, recognition or a license check contacts a server, the tool cannot run in a sealed environment.
6. Ask for approval before using developer services. Disconnected use of a cloud-hosted product may need written approval first.

Yap, which we make, is an on-device dictation app for Mac and Windows. Treat it like any other candidate: install it while connected, move the files across, and run the test above in each language before you approve it.

For the wider picture, read the guide to dictation at work and the overview of voice privacy. Related questions: does Windows voice typing work offline, is offline speech recognition more private than cloud, and can Whisper AI run offline.