1. Voice typing (Win+H) stops when the PC is offline
Microsoft’s voice typing page is explicit: voice typing uses online speech recognition powered by Azure Speech services, and to use it you need an internet connection, a working microphone, and the cursor in a text box. Press Windows key + H on a hardware keyboard, wait for the “Listening...” alert, and say “Stop listening” or select the microphone to end. Windows key + Alt + H moves through the voice typing menu from the keyboard. The page applies to Windows 11 and Windows 10.
Copilot+ PCs add Fluid dictation, which corrects grammar, punctuation, and filler words with small language models that run on the device. Those models install in the background, and the option stays dimmed in voice typing settings until the download completes. The correction step is local, but Microsoft still describes the recognition itself as online, so plan for Win+H to fail when the PC is disconnected, Copilot+ or not.
Voice typing is still the right first tool for occasional dictation on a connected PC. The gear icon on its flyout controls the launcher, automatic punctuation, the profanity filter, the default microphone, and how long it waits before acting on a voice command.
2. Voice access is Microsoft’s offline dictation on Windows 11
Microsoft describes voice access as a Windows 11 feature for controlling the PC and authoring text using only your voice and without an internet connection. It requires Windows 11 version 22H2 or later. The connection matters once: the first time you open voice access you choose a language and select Agree and continue, and it downloads the language files used by on-device speech recognition. If you exit before that download finishes, setup starts over the next time.
Turn it on under Settings → Accessibility → Speech → Voice access, or search for “voice access” and select Open. Checkboxes on the same page start it automatically before or after you sign in. After the model downloads, setup asks you to confirm a default microphone; once one is selected you can start dictating, and the voice access bar stays at the top of the main screen.
The limits are language and scope. Microsoft’s setup page lists English (US, UK, India, New Zealand, Canada, Australia), Spanish (Spain, Mexico), German, French (France, Canada), Simplified Chinese, Traditional Chinese, Japanese, and Italian. Microsoft documents voice access for Windows 11 only, so Windows 10 PCs do not have it, and because it is a voice-control system as well as a dictation tool, it listens for commands alongside your text.
- Internet is needed only to download the language files during setup.
- Microphone use follows Settings → Privacy & security → Microphone, including desktop-app access.
- Change the language later under Voice access settings → Languages.
3. Windows Speech Recognition and Microsoft 365 Dictate
Windows Speech Recognition (WSR), the older command-and-dictation tool, is on Microsoft’s list of deprecated Windows features in an entry dated December 2023: it is no longer being developed and is being replaced by voice access. It may still be present on older installations, but a new offline workflow should not depend on a feature Microsoft has stopped developing.
Dictate in Word and the other Microsoft 365 apps is a connected service. Microsoft says it needs a microphone and a reliable internet connection, is available only to Microsoft 365 subscribers, and sends your speech to Microsoft only to return text; the same page says the service does not store your audio or the transcribed text. It suits long documents on a connected PC with a subscription, not dictation on a plane or an isolated network.
4. Offline dictation apps for Windows
Third-party apps fill the gap between voice access and voice typing: a global hotkey, a speech model stored on the PC, and text placed at the cursor in whatever app has focus. Every one of them has to download a model before its first offline dictation, so do that while you are still connected.
SpeechPulse runs Whisper recognition on Windows and macOS and supports offline speech recognition, so voice and text stay on the machine. It types into any text field, transcribes 99 languages, offers push-to-talk or automatic speech detection, and can transcribe audio files with speaker diarization; a 30-day free trial comes before purchase. Its optional AI templates call language-model APIs, so leave them off when offline is the requirement. Handy is a free, MIT-licensed app for Windows, macOS, and Linux that transcribes locally with Whisper or Parakeet models; it is also published on winget as cjpais.Handy, a package the Handy developers note they do not maintain.
The model you pick sets the download and the load on the PC. Handy’s manual-install list gives a sense of scale: Whisper Small is 487 MB, Whisper Large (quantized) 1,100 MB, Whisper Turbo 1,600 MB, and Parakeet Unified EN 0.6B 731 MB. Handy also documents copying model files in by hand for PCs behind a proxy or firewall, which helps when the machine that will dictate offline cannot download models itself.
5. Prove offline behavior with an Airplane mode test
An offline label describes the speech step, not everything the app does. Install the app, activate it if it requires that, and download the model you intend to use while connected. Then switch on Airplane mode or unplug the network cable, quit and reopen the app, and dictate the same paragraph into Notepad, a browser text box, and the program you actually work in.
Watch for three failures: recognition that stops with a network error, a sign-in or license screen that appears after the restart, and a feature that quietly switches to a cloud model. Also note how long a paid license tolerates being offline. Periodic license checks are normal, but the grace period decides whether a travel laptop keeps working for a week or for a month.
- Download the model on a connected network first.
- Restart the app while disconnected, not just the Wi-Fi.
- Test the app you really dictate into, not only a demo field.
- Write down the license’s offline grace period.
6. IraVoice on Windows: offline after one model download
IraVoice 1.2.2 runs on Windows 10 or 11, 64-bit (x64). Speech recognition runs on the PC with NVIDIA Parakeet: the default Parakeet Unified EN model, about 700 MB, downloads once on first run, and from then on dictation works with no connection. Optional offline models include Parakeet v3 for 25 European languages and Whisper models for many more, which are heavier.
Hold Right Ctrl, speak, and release: the text is pasted with Ctrl+V into the focused app, and your clipboard is restored after that app has read it. Tap Right Ctrl instead for hands-free dictation and tap again to stop; Esc cancels. IraVoice does not press Enter, so you review the text first. When the focused app runs as administrator, Windows blocks synthetic input from a normal app, so IraVoice leaves the text on the clipboard and says so.
What still connects: the one-time model download, plus trial and license check-ins and update checks to api.iravoice.com that carry a salted hash of a hardware ID, never your audio or text. The license checks in weekly and keeps working for 30 days without a connection. IraVoice Local costs $79 once for one computer, $109 for two, or $149 for three, after a 3-day free trial that starts with the first dictation and needs no email, card, or account.
- Default key: Right Ctrl, changeable under General.
- Other paste methods: Shift+Insert, Ctrl+Shift+V, direct typing, or none.
- In Windows Terminal, PowerShell, cmd, VS Code, Cursor, Visual Studio, and JetBrains IDEs it spells names from your Git project, such as useEffect.