The airplane test is the most honest benchmark in this category, because you can't argue with it. Turn off Wi-Fi and every dictation tool reveals its architecture: the on-device ones keep typing, and the cloud ones become a microphone connected to an error message. This guide is about ending up in the first group — on purpose, with the setup done before you're at 35,000 feet.
Why bother dictating offline?
- Planes, trains, and dead zones. The cliché is a cliché because it happens. Long-haul flights are where half the world's overdue writing finally gets done — and hotel Wi-Fi that technically exists is its own dead zone.
- Privacy that doesn't depend on a policy. A privacy policy says a company intends not to misuse your audio. Offline architecture means the audio physically cannot leave. One of these survives an acquisition; the other is a paragraph that can be edited. We've written down what leaves your Mac and when — the offline answer is "nothing".
- Compliance without meetings. If you handle patient notes, privileged drafts, or anything under NDA, "the audio never crosses the network" turns a data-processing review into a non-event.
- It keeps working. No outage page, no degraded-service banner, no "we're experiencing elevated error rates" during your deadline. Local software has boring uptime.
What "offline" actually means, tool by tool
Marketing has muddied "on-device" badly enough that it's worth being precise:
- Cloud-transcription services (Wispr Flow and similar, per their docs as of July 2026) send audio to servers. Offline, they stop. That's not a flaw so much as the deal you accepted — the cloud is where their product lives.
- Apple's built-in dictation processes many languages on-device on Apple Silicon, so it largely works offline — with the usual built-in limitations (silence timeouts, no custom vocabulary, raw output) that we covered in the how-to-dictate guide.
- Whisper-based apps — OpenVoiceFlow among them — run the speech model on your Mac's own silicon. Transcription is offline by construction, not as a fallback mode with reduced accuracy. The full model runs locally; there is no lesser "offline version" of it.
One asterisk applies to every tool in the last group: the model itself has to be downloaded once, and that requires internet. Do it at home, not at the gate.
The fully offline pipeline, step by step
Here's the setup we'd do on a new Mac, in the fifteen minutes before leaving for the airport:
- Install OpenVoiceFlow. Grab the DMG, drag to Applications, run the one-minute setup — permissions, then a Whisper model download. Pick the model while you have bandwidth; our model guide says
large-v3-turboon Apple Silicon,smallon Intel. - Confirm you're already done. Cleanup ships Off, which means the default pipeline — hold key, speak, Whisper transcribes, text pastes — is already 100% local. Turn Wi-Fi off and try it. That's the whole test.
- Optional: local AI cleanup with Ollama. If you want grammar fixes and filler-word removal without touching a cloud, install Ollama and pull a small model —
ollama pull llama3.2gets you a 3B-class model that handles cleanup comfortably on Apple Silicon. Then pick Ollama as the cleanup backend in OpenVoiceFlow's settings. Transcript goes tolocalhost, polished text comes back, nothing crosses the network. - Teach it your words. Two minutes adding names and jargon to the personal dictionary saves you from fixing the same misspelling on every flight for the rest of the year.
From here on, the airplane test passes every time. Audio in, polished text out, radios off.
What you give up offline (and what you don't)
Fairness requires the other column. Fully offline, you forgo cloud cleanup models — the frontier LLMs you'd reach through OpenRouter, OpenAI, Anthropic, or Groq are better writers than a local 3B model, and it isn't close. For dictation cleanup, though, the job is deleting "um", fixing commas, and joining fragments — a job a small local model does perfectly well. You're not asking it to compose; you're asking it to tidy.
What you don't give up is the thing that matters: transcription quality. The Whisper model on your Mac is the same model, same weights, connection or not. Accuracy offline is accuracy, full stop.
And the hybrid answer is always available: cleanup Off or Ollama on the plane, your OpenRouter key when you land. It's a dropdown, not a commitment.
FAQ
Can I dictate on a Mac in airplane mode?
Yes — with on-device transcription. OpenVoiceFlow works identically with radios off; the transcription hardware is your Mac itself.
Is offline dictation less accurate?
No. The full Whisper model runs locally — there's no degraded offline mode. Your model choice and mic matter; your connection doesn't.
Can the AI cleanup step really run offline too?
Yes, via Ollama on the same machine. Small instruct models handle grammar-and-filler cleanup well. Cleanup is Off by default anyway, so the out-of-box pipeline is already fully local.
Does OpenVoiceFlow phone home at all?
The app has no telemetry and no account system. Auto-update checks fetch a signed feed from this website when online; audio and transcripts are never uploaded, as the privacy page spells out — and the source is public if you'd rather verify than trust.
Set it up before the flight
Free, open source, and indifferent to your Wi-Fi. Fifteen minutes now, fully offline dictation forever.