Speech to text offline: how to keep dictating when the internet goes

Offline means two jobs, not one

Dictation is two steps. Recognition turns your voice into words. Cleanup turns those words into a sentence you would send: punctuation, structure, the "um"s taken out. A tool can do one on your device and the other on a server, and "works offline" can describe only the first.

In Deixis, recognition always happens on the device, on every platform, and the network is only for optional extras such as cleanup by a language model, translation and meeting notes. Cleanup itself starts on your machine, with a rules engine that punctuates and tidies each sentence. A language model can take over for harder cases, such as a transform you ask for ("make this a bullet list") or a long stretch with no punctuation, but only if you set one up. With no cleanup endpoint set, dictation uses the local rules and no text leaves your machine.

If you have set up a cloud endpoint and the network drops, dictation keeps working. A sentence that cannot reach the model gets the rules output instead, the same thing that happens when the endpoint is slow to answer.

Before you go offline

A few minutes with a connection saves a frustrating morning later.

1. Download the speech model. On Windows the installer does not carry it, so the first launch shows a single Download button: 142 MB, fetched once. Until it finishes there is no speech model, and dictation says so rather than failing quietly.
2. Choose your model size now. On Windows, Settings → Speech model offers the standard model and a larger, more accurate one (465 MB) that downloads when you switch to it. If you want the larger one while you are away, switch before you leave.
3. Grant permissions on a Mac. macOS asks for Accessibility and microphone permission the first time you dictate, and you grant both in System Settings → Privacy & Security. Do it once at your desk.
4. Decide what cleanup should do. If you use a cloud endpoint you can leave it alone, since sentences fall back to rules. If you would rather dictation did not try the network at all, set Settings → Cleanup to Never on Windows; translation and meeting notes still use the endpoint when you ask for them.
5. Test it. Switch off the wifi and dictate a sentence into something you use every day.

What works with no network, device by device

Windows has the full set. With the network gone you still have dictation at the cursor in any app; free mode, where you double-tap the chord and narrate while collecting screen grabs and selections into one note; the dictionary and its vocabulary packs; screen grabs and the annotation editor; and screen recording, with what you said transcribed on the machine when a microphone was recorded. Meetings work too: Deixis records your microphone and, where the system allows, the other side of the call, transcribes both on the machine, and writes the transcript with no key, no account and no network.

Mac runs hold-to-dictate at the cursor, free mode and its toolbar, screen grab and annotation, and screen recording, with speech recognised on the Mac itself. Meeting mode is not in that list.

Android puts a floating microphone bubble over any app's text field, and records meetings from the phone's microphone, transcribing them on the phone as they run. The dictionary and both speech models are there too. It is on Google Play.

iPhone gets a keyboard: switch to it, hold the microphone key, speak, and the text lands in the field. The speech model runs inside the Deixis app on the phone, and the keyboard only inserts the result. It is in beta on TestFlight, not yet in the App Store, and meetings are not on the iPhone yet. The platforms page has the current state of each.

What needs a network, and how to keep it local anyway

A handful of features need a language model to answer. With no model configured, or no network to reach one, this is what each does:

• Cleanup falls back to the local rules, as above.
• Translate needs a model, so it does not run.
• Meeting notes wait. The meeting still records and its transcript is still written, because notes are deliberately the last step, and Regenerate notes retries them later.
• A meeting's name comes from the first substantial sentence somebody said when no model is configured, instead of a model's short phrase. A name you type yourself is never overwritten.
• A screen recording's title and summary wait too. You still get the transcript and the chapters, and the Library row says why there is no summary and offers Summarise for later.

On Windows you can point all of these at a model on your own hardware instead, and then none of them needs the internet. Run a local server such as Ollama, pull a model such as qwen3:8b, and point Deixis at it in Settings → Cleanup. A local model wins over any cloud endpoint rather than falling back to it, and if the local server is down there is deliberately no fallback to the network: cleanup uses the rules and dictation keeps working. Nothing is metered, because the work happens on your machine. Settings, and where cleanup runs has the details.

Two features cannot be made local. Share a link uploads a grab, note or screen recording you picked so somebody else can open it, which needs a network and a Deixis account; and the Claude usage key, when you press it, asks Anthropic for your plan's usage. The two calls Deixis makes on its own on Windows and the Mac, the update check at launch and a check for an announcement when you open Settings, are not needed for dictation to work.

How to check another tool before you rely on it

Before a trip, switch off the network and dictate. If nothing appears, the tool most likely recognises speech on a server. Windows' own voice typing is in that group: Microsoft's support page said, when we checked on 5 September 2026, that it uses online speech recognition and needs an internet connection.

Then check the second step. A tool can recognise locally and still send every sentence to a cloud model to be tidied, which means the words stop arriving tidy, or stop arriving at all, when the connection goes. And check whether it needs you to sign in before it starts. Deixis never asks you to sign in to dictate, and dictation is free with no account.