Offline Speech to Text on Mac: What Actually Runs Locally

"Offline" and "private" get used interchangeably in dictation marketing, and they are not the same claim. Offline means the transcription happens on your hardware. Private means the audio does not end up on someone else's. A tool can be neither, one, or both — and the only reliable way to find out is to disconnect and see what still works.

Here is where the common options actually stand.

What genuinely runs offline#

Apple's built-in dictation, on Apple silicon. On a Mac with Apple silicon, general text dictation — composing a message, writing a note — is processed on the device in many languages, and no internet connection is required. This is Apple's documented behaviour, not an inference. Two carve-outs matter: dictating into a search box uses server-based dictation, and languages or hardware without on-device support fall back to Apple's servers. If you are on an Intel Mac, assume server dictation unless you have tested otherwise.

You can check what your own machine is doing in System Settings, Keyboard, Dictation — Apple surfaces there whether your voice input is processed on-device rather than sent to Siri servers. See Apple's dictation guide.

Voice Control. This is the accessibility feature for driving the whole Mac by voice, not just typing. When you turn it on for the first time your Mac needs a connection for a one-time download; after that, no internet connection is needed to use it. It is the most capable fully-offline voice feature Apple ships, and most people have never opened it.

Local Whisper-derived apps. A category rather than a product: speech models small enough to run on a laptop, packaged as Mac apps. The transcription happens in memory on your machine, so the network state is irrelevant. This is the approach Jarvis takes — the model runs on your Mac, audio never leaves it, and text comes back in around 240ms because there is no server round-trip to wait for.

Two commercial ones belong here rather than in the cloud list. Superwhisper has the clearest disclosure in the category: it splits the work into a voice stage and a language stage and lets you set each independently to local or cloud, and says of the local path that "no audio or text ever leaves your machine." Its own caveat is worth repeating — offline models "only run really well on Apple Silicon macs," so on Intel this is effectively a cloud product. MacWhisper transcribes locally by default, and is built for files rather than live dictation, which is a different job again.

What does not#

Google Docs voice typing. Google's own documentation scopes it to Docs and to speaker notes and captions in Slides, on recent Chrome, Edge and Safari. It is a browser speech service, so it needs a connection, and it does not exist outside those two products — not in Gmail, not in a text field anywhere else. See Google's page.

Microsoft 365 Dictate. Microsoft states plainly that dictation requires a microphone and a reliable internet connection, and that it is available to Microsoft 365 subscribers. It is a cloud service with a button inside Word and Outlook. Details on Microsoft's support page.

Wispr Flow. The most-recommended paid dictation app on the Mac is cloud-only, and to its credit it says so in its own documentation rather than burying it:

Transcription always happens in the cloud to provide the best speed and accuracy.

Its Privacy Mode is worth understanding precisely, because the name invites a wrong reading. It controls retention, not location — with it on, Wispr keeps nothing; with it off, audio is discarded after each request. Either way the audio was transcribed on a server. There is no offline mode. wisprflow.ai/privacy

Willow Voice is the one to look at carefully, because its own properties disagree. The homepage feature list advertises "Offline mode — Dictate without an internet connection, powered entirely by your device." Its help centre, under the heading Why We Use Cloud Models, says the opposite: "Willow uses cloud-based AI models" and "On-device models cannot match this level of performance right now," with no mention of an offline mode anywhere on the page. Both cannot be true as written. If offline is your requirement, ask them directly which plans include it and which model powers it before paying. help.willowvoice.com

A monthly price remains a reasonable tell in general: on-device inference costs the vendor nothing per minute, so a per-seat subscription usually means somebody is paying for GPUs.

Test yours in thirty seconds#

1. Turn Wi-Fi off. Actually off, not "disconnect from this network" — and unplug Ethernet if you use it.
2. Open a plain text field. TextEdit or Notes, not a search bar, not a browser.
3. Dictate two sentences.
4. Watch the failure mode. Instant text means local. A spinner, a delay, or silence means it was going somewhere.
5. Repeat in a search field. If that one fails while the text field works, you have confirmed the server carve-out on your own machine.

What you give up, honestly#

Running locally is not free of trade-offs.

• Disk and first-run download. Local models are files. Expect to fetch one before your first offline session.
• Hardware dependence. Apple silicon is where local speech models are comfortable. Older Intel hardware works but with less headroom.
• No server-side context. Cloud services can lean on much larger models. In practice, for conversational speech, the gap is far smaller than it was two years ago — but a local model is a fixed thing on your disk, not one that quietly improves each month.

What you get in return is that the network stops being a dependency. No spinner on hotel Wi-Fi, no degradation on a train, no transcript of your working day sitting in a vendor's logs — and for anyone under a confidentiality obligation, that last one is often the whole argument.

If that is your requirement, Jarvis is built for it: free, open source, Mac-only (Apple silicon and Intel, macOS 12 or later), on-device dictation with no phrase limit, working in any app you can type into. 2.0 is in beta. There is a broader comparison in the best AI voice assistant for Mac, and if your current setup is misbehaving, start with Mac dictation not working.