How to dictate on a Mac without sending your voice to the cloud
What on-device really means, how to check whether your audio is leaving your Mac, and the ways to dictate offline on macOS 26.
· 6 min read
Yes — you can dictate on a Mac with the Wi-Fi switched off, and on macOS 26 you can do it well. The catch is that “on-device” and “private” get used to mean several different things, and an app can be telling the truth about one of them while your audio is still going somewhere.
This is how to tell the difference, in an order you can actually work through.
“On-device” means three separate things
A dictation app does at least three jobs, and each one can independently run on your Mac or on somebody's server:
- Transcription. Audio in, raw words out. This is the one everybody advertises.
- The rewrite. Raw words in, clean sentence out — fillers removed, punctuation added, “Thursday, no, Wednesday” resolved. This is usually a language model, and a language model is the easiest of the three to quietly put in a data centre, because it is the most expensive to run locally.
- Everything else. Sync, history, account, licence checks, crash reports, analytics.
An app can transcribe locally and still send the transcript out to be tidied up. That is not fraud and it is often disclosed — but it does mean the sentence you spoke left your Mac, which is exactly the thing you were trying to avoid. When you read a privacy page, read it for all three jobs, not the first one.
The clearest possible version of that disclosure is worth quoting, because it shows what an honest cloud answer looks like: Wispr Flow's own data-controls page says “Transcription always occurs on the cloud. This is the best way for us to provide accurate, low latency transcription.”3 No ambiguity, and a perfectly defensible engineering choice — just a different one to local.
How to check where your audio actually goes
- Search the privacy policy for the word “audio”. Not “data”, not “privacy”. “We never sell your data” is a statement about sales, not about receipt.
- Ask macOS. Apple documents that you can check in Keyboard settings whether your voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers.1 That is the built-in feature's own answer, from Apple, in the settings pane.
- Pull the plug. Turn Wi-Fi off and dictate a long paragraph. Nothing beats this test: if clean text still lands at your cursor, both the transcription and the cleanup are running locally. If it fails, or if it produces raw text where it normally produces tidy text, you have just found which stage needed the network.
- Watch the connections. A per-app firewall (Little Snitch, LuLu) or Activity Monitor's Network tab shows what an app talks to. Expect some traffic even from a local app — licence activation, update checks — and judge it by volume and timing: a connection every time you speak is not an update check.
- Look for a model download, but do not use size as proof. Whisper and Parakeet builds ship real model files: for reference, WrapScribe's optional Parakeet download is about 480 MB and Whisper Large v3 Turbo is 1.6 GB. But an app that uses Apple's own on-device speech models can be tiny — WrapScribe's installer is around 15 MB — because the models belong to macOS, not to the app. A small download is not evidence of a server.
The three ways to dictate offline on macOS 26
1. Built-in Dictation
Free, already installed, no permissions beyond the microphone. Apple documents no length limit — “You can dictate text of any length without a timeout” — but it stops when no speech is detected for 30 seconds, which is why long thinking pauses end a session.1 Availability varies by language and region, so check yours rather than assuming. What you get is a transcript: no cleanup, no vocabulary of your own, no rewriting. Where that is and isn't enough is its own comparison.
2. An app that runs Whisper or Parakeet locally
The open-source route. You download a model, it lives on your disk, transcription happens there. The trade is disk space and, for the larger models, speed — and the reward is language coverage: Whisper handles around 99 languages, which is far past what any built-in option covers. Check the cleanup question separately, because some apps in this category pair local transcription with a hosted language model.
3. An app built on Apple's on-device stack
macOS 26 (Tahoe) introduced SpeechAnalyzer, Apple's on-device speech framework, alongside on-device Foundation Models. That combination is what lets an app do both jobs — transcribe and rewrite — without a server. It is fast: MacStories transcribed a 34-minute, 7 GB video in 45 seconds with the new APIs, against 1 minute 41 seconds for MacWhisper's Large V3 Turbo, with no noticeable difference in quality.2 The cost is the requirement: Tahoe, on Apple silicon.
What offline does not buy you
On-device is not encryption. If your transcripts live in a database in your home folder, anyone with your unlocked Mac can read them. Turn FileVault on. Find out where the app keeps its history and whether you can delete it.
Offline does not mean no account. Most paid apps still need one connection to activate a licence, and some sync preferences. That is fine — the question is what the sync payload contains. Preferences and counters are not sentences.
Offline does not mean no telemetry. Analytics on a website is a different question from analytics in an app; check both.
Where WrapScribe stands, precisely
So the claim can be checked rather than believed: WrapScribe transcribes with Apple's on-device speech model, and runs Smart Cleanup and meeting notes through Apple's on-device Foundation Models. There is no endpoint on our side that accepts audio or transcripts — not a policy we could quietly reverse, an API that has nowhere to put them. Recorded audio, transcripts, history, dictionary, snippets and meeting notes stay on the Mac. What does sync to an account is an email address, a name for each activated Mac, which licence key was redeemed, per-day word and dictation counts, and your preferences. Once a Mac is activated and a language model is downloaded, everything works with no connection at all.
It needs macOS 26 (Tahoe) on Apple silicon, it is free to download, and a key is $100 for one Mac, $200 for three, $500 for every Mac you own — one-time, not a subscription. If you are on an older macOS or an Intel Mac, option 1 or option 2 above is your answer, and that is a genuine answer rather than a brush-off.
Does Apple's built-in Dictation work offline?
Apple documents that you can check in Keyboard settings whether your voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers. Because availability varies by language and region, the honest answer is to look at that setting on your own Mac in your own language rather than take a general claim.
Is on-device dictation less accurate than cloud dictation?
Not necessarily, and the gap has closed. In one published hands-on test, MacStories transcribed a 34-minute video with Apple's new on-device speech APIs in 45 seconds against 1 minute 41 seconds for MacWhisper's Large V3 Turbo model, and reported no noticeable difference in transcription quality. That is one file on one machine, not a benchmark suite, so treat it as an indication rather than a measurement of your own audio.
Does WrapScribe need an internet connection?
Only to activate a licence key and to download a speech model. After that, dictation, Smart Cleanup and meeting notes all run with no connection at all, because they run on Apple's on-device models rather than on a server.
Does “on-device” mean my dictation is encrypted?
No. On-device means the audio does not leave your Mac. It says nothing about what happens to the transcript stored on that Mac. Turn FileVault on, and check where any dictation app keeps its history and whether you can delete it.
Sources
- Apple, “Use dictation on Mac”, macOS User Guide — support.apple.com/guide/mac-help/use-dictation-mh40584/mac. Source of the on-device check in Keyboard settings, the “any length without a timeout” wording, and the 30-second silence rule. Fetched 19 August 2026.
- John Voorhees, “Hands-On: How Apple's New Speech APIs Outpace Whisper for Lightning-Fast Transcription”, MacStories, June 2025 — macstories.net. Source of the 45-second and 1:41 timings and the quality observation. Fetched 19 August 2026.
- Wispr Flow, “Data Controls” — wisprflow.ai/data-controls. Source of the quoted sentence about cloud transcription. Quoted as documented on 19 August 2026; vendors change their architecture and their pages, so check it yourself before relying on it.
- WrapScribe's own architecture, model sizes, requirements and prices come from this site — the privacy page and the download page.