Offline dictation on Windows

What "offline" actually means

The word gets stretched. A tool can call itself offline because your notes are stored on your disk, or because it caches text while the connection is down and uploads it later. None of that is the thing that matters. There is exactly one question worth asking: where does the audio turn into words?

If recognition happens on your device, the audio never has to go anywhere, and dictation keeps working when nothing else does. If recognition happens in the cloud, your voice is a network request — which means no connection, no words, and a copy of what you said sitting on somebody's server for however long their policy says. Everything else about a dictation tool is a detail by comparison.

There are three honest answers, and the table further down gives each tool's:

• On your device. The engine and the model sit on your machine. Dictation works with the network unplugged. This is where Deixis is, always, on every platform it runs on.
• Either, your choice. The tool ships a local engine and a cloud one, and a setting picks. These are genuinely fine choices — just check which way the setting ships, because the default is what most people will actually use.
• In the cloud only. No connection, no dictation. Windows' own voice typing — the Win+H feature — is in this group: Microsoft's support page says it uses online speech recognition and needs an internet connection. (Microsoft's separate Voice access feature does work without one, in a few languages, but it is a way of controlling the computer rather than the dictation feature people mean when they ask.)

A second question is worth asking straight after: does it need an account? A tool can recognise on your device and still refuse to start until you have signed in, which puts a network dependency back in front of the thing you installed to avoid one. Deixis has never asked anyone to sign in to dictate, and it never will.

What Deixis does when the network is gone

Nearly all of it. With no cleanup destination configured, these all run entirely on your machine, with no account and nothing to switch on:

• Dictation. Hold the chord, speak, release, and the text lands at the caret in whatever app has focus — your editor, your terminal, your email, a browser box. Your clipboard is handed back the way you left it.
• Free mode. Double-tap the chord instead and it keeps listening until you double-tap again, while you drag a region of the screen, pick an element or take a selection. What you said and what you pointed at come out as one note, pasted at the caret and saved as a file.
• Your dictionary. The single biggest accuracy lever, and it costs nothing in speed. Select a word anywhere, press one key, and it is in your vocabulary; names it keeps mishearing can be corrected once and stay corrected. Five vocabulary packs ship in the box and you can drop in your own.
• Screen grabs. Freeze the screen, drag a region, quick-save it or open the annotation editor — arrows, boxes, pen, highlighter, text, redaction and numbered step badges. Every save keeps a sidecar so the grab stays re-editable rather than flattened forever.
• Screen recording. One chord records the screen you are on to a video file and the same chord stops it, with a small bar and the tray icon showing that it is rolling. If a microphone was recorded, what you said is transcribed on this machine. The short title and summary written for each recording are the one part that wants a language model: point Deixis at one running on your own machine and that stays offline too, and with nothing configured you still get the transcript.
• Meetings. Deixis records your microphone and, where the system allows it, the other side of the call, and transcribes both on this machine while they run. The readable transcript is written to disk with no key, no account and no network at all. It tells the remote voices apart and remembers the ones you name, and the voiceprint that does it stays on the machine, is never uploaded, and can be deleted per person.
• Your own buttons and workflows. Add an action to the palette that runs a program with your finished note and types back whatever it prints — file the ticket, get the key back at your caret. Or chain steps together as plain settings, with a template library of ready-made ones. Webhooks go to URLs you own, signed, and never through anybody else's cloud.

Now the honest part, because a privacy page is only worth reading if it says what does leave.

Two calls a fresh install makes on its own, and neither carries anything you say or type. The update check runs at launch and when you ask for it in Settings; the version comparison happens on your machine. At launch it carries the app's version, a once-a-day mark that this copy was opened and, once each, the first steps a new copy reaches, with no ID, so we know how many people use Deixis; switch off Settings › Privacy › Count this copy and it carries nothing but the request. And when you open the Settings window, Deixis asks whether we have posted an announcement — a short notice, drawn as a popup you can dismiss. Which ones you have dismissed stays on your machine, and announcements have an off switch of their own.

The speech model is a one-time download you press a button for. The installer does not carry the weights, so the first launch shows one Download button. That request carries nothing but the request itself. After it finishes, dictation is done with the network.

One key asks a question of somebody else's service, and only when you press it. Ctrl+Win+C shows how much of your Claude plan is left, which it can do only by reading the sign-in Claude Code has already stored on this machine and asking Anthropic. It needs no endpoint of yours because neither the sign-in nor the address it goes to is ours, and it carries nothing of yours besides that sign-in: no audio, no text, no transcript. Never press it and nothing ever goes to Anthropic.

Three optional routes exist and are off until you configure a destination: cloud cleanup of a dictated sentence, translating a selection, and meeting notes, along with the short titles written for meetings and recordings. With no destination set, cleanup falls back to a rules engine on your machine, a meeting is named from its first real sentence, and the rest simply do not happen. And if you want a language model's polish without a network, you can point all of them at a model running on your own machine instead — that is a supported setting, not a workaround, and there is deliberately no fallback to the internet if that local server is down.

Once the speech model is in, no setting uploads your audio, on any plan. The recording a dictation makes and the audio of a meeting stay on the machine, and so does a screen recording's narration unless you share that recording, below. There is no setting that sends audio out to be recognised, and no feature that depends on one. Before it is, a first run may choose to dictate on Deixis's cloud while the model downloads.

The one feature that uploads a file of yours is handing a recording, a grab or a note to somebody as a link, because being uploaded is what a share link is. It is a file you picked, at the moment you pressed Share; a meeting recording cannot be shared at all. Making a share link needs a Deixis account, though the person you send it to needs nothing, no app and no sign-in. Everything that produced the recording happened offline; only the sharing does not.

How the tools compare

Deixis is at the top because this is the Deixis website and Deixis is our product — the rest is alphabetical, and the facts in every row come from that maker's own pages on the date beside it. Read down the middle column: it is the one that decides whether the tool is any use to you on a plane. Ours and Whisperstream's are the two rows that answer both columns without a condition attached.

The fair thing to say is that Deixis is not alone. Whisperstream is the closest tool on this site — push-to-talk dictation on Windows, recognition on your own processor, no account — and it is a paid product after a short trial, where Deixis dictates free forever. Dragon Professional has been doing local recognition on Windows since long before any of this, and remains the serious choice for people who dictate all day and need its voice commands and macros. Talon runs its recognition locally too, and is aimed at controlling the whole computer by voice rather than at typing prose — for people who cannot use a keyboard at all, it is a different and often better answer.

Four more give you both a local engine and a cloud one. Superwhisper transcribes with a local Whisper model on Windows as well as on the Mac, and on Pro can run its AI modes on a language model on your own PC too; Willow Voice lists offline dictation on its paid individual plan without saying which platforms carry it; and VoiceInk and MacWhisper, both Mac software, recognise on the Mac by default and reach a cloud service only if you connect one under your own key.

Where Deixis differs from all of them is the surface area that stays offline. On Windows it is not only the dictation: free mode, the grabs and the annotation editor, the screen recordings with their local transcripts, the meeting recording and transcription, the dictionary and your own buttons all run with the network unplugged. The other three builds carry a different slice each — the Mac has dictation, free mode, grabs and screen recording; the iPhone is a keyboard, in beta on TestFlight; Android is a dictation bubble with meetings — and what they have in common is the part that matters here: recognition happens on the device, every time.

Why offline dictation is worth choosing

The practical case is the easy one. A plane, a train through a tunnel, a hotel with a captive portal, a client site whose guest network blocks everything interesting, a machine that has no route out at all — these are ordinary working conditions, and they are exactly when a cloud dictation tool becomes a button that spins. A tool that recognises on your device does not have good days and bad days. Its speed does not depend on somebody else's servers, and it does not stop working because a provider is having an incident.

The privacy case is quieter and, for a lot of people, the one that decides it. Your voice is not the same as your text. It carries who you are, who is in the room, and everything you said before you thought better of it. If it never leaves your machine, no policy has to be trusted, no retention setting has to be checked, and no account has to be kept honest. That is not a claim about anybody else's conduct — it is that the safest data is the data that was never sent.

It also makes an easier conversation at work: "the audio stays on the laptop" is a sentence a security reviewer can check by unplugging the network.