Is Mac dictation private? Where your voice goes.
Mac dictation is private when your speech is turned into text on your Mac and no audio is uploaded. That depends on the app, its settings and your language. Apple’s own Dictation can go either way, some apps always use a server, and clean-up, summaries and usage data can leave your Mac even when the audio stays.
This is for anyone who dictates client notes, drafts or code and wants to know what leaves the machine. You get each app’s own description, what we checked in Noted’s code, and a 20-minute test you can run tonight.
TL;DR
- Apple Dictation can run on your Mac or send audio to Siri servers; Wispr Flow says it always transcribes in the cloud; Superwhisper, OpenWhispr and Noted can keep speech to text on your Mac.
- Audio is one leg of three. Clean-up and summaries, and usage data, can each go somewhere else.
- You can check: turn Wi-Fi off, then watch the app’s connections with
nettop. - Our founder dictated in Noted with Wi-Fi off, by hand, on Mac and Windows, November 2026. That shows dictation
works with no connection, not what leaves the computer: we have not run the
nettopcheck on any app. Rival facts come from their own pages, read 12 November 2026. What we did check is in “What we have and have not tested”.
Where does my voice go in each app?
Dictation has three legs: speech to text, clean-up or summaries of that text, and usage data. Each can run on your Mac or on someone’s server, so ask where each one goes. This is each app’s own description, from its help or privacy page.
| App | Speech to text | What else can leave your Mac | Read |
|---|---|---|---|
| Apple Dictation | System Settings › Keyboard › Dictation says whether general text dictation is processed on your Mac or sent to Siri servers (not search boxes). | If you opt in to sharing audio recordings, Apple can store audio of your Siri and Dictation use. Switch it off in Privacy & Security › Analytics & Improvements. | 12 Nov 2026 |
| Wispr Flow | Always in the cloud, per its data usage page. | Usage details such as word count, app and timestamp are sent even with sharing off. Cloud storage of dictation history and the model-improvement toggle both default to on, unless restricted, on its help page. | 12 Nov 2026 |
| Superwhisper | Local models: audio stays on your Mac. Cloud models: audio goes to a voice provider. | With cloud models, transcribed text goes to a language-model provider too. Share links upload the summary. | 12 Nov 2026 |
| OpenWhispr | Local mode: audio stays on your device. Cloud mode: sent to its servers, then a model provider. | Cloud mode also sends your dictionary terms. Usage analytics and cloud copies of notes are off by default. | 12 Nov 2026 |
| Noted | Runs on your Mac. Audio is never sent to us or anyone else. | Cloud features send text (see below), and the app sends usage analytics unless you turn them off. | 12 Nov 2026 |
Where each rival is better. Apple’s is built in, costs nothing extra and needs no new account. Wispr Flow’s help page documents its data controls in detail and says its AI providers are barred from training on your data. Superwhisper can keep both speech to text and the language-model step on your Mac, with a “device only” setup it documents. OpenWhispr lets you pick local or cloud separately for dictation, notes and uploads.
Where Noted is better. Speech to text has no cloud option at all, so no setting can send your audio by mistake, where Wispr Flow says it always transcribes in the cloud (data usage; see also Noted compared with Wispr Flow). Your notes and dictation history stay on your computer, with no cloud copy unless you share a note by link. And one Cloud switch turns every text feature off.
What does Noted send?
From our privacy page, last updated 23 October 2026: speech is turned into text on your computer and audio is never sent to us or anyone else. Your notes, transcripts and dictation history stay on your computer, so plain dictation involves no voice upload.
The rest is text, and only with Cloud switched on. Polish sends the text you polish, with your dictionary words. AI meeting notes, Ask, Live Answers and screen answers each send what they need when they run, to AI providers in the United States, to be processed and not stored. One Cloud switch in Settings › General turns all of that off. Noted also sends usage analytics (app opens, features used, counts and timings, never your text) to PostHog in the EU; Settings › Privacy has a switch for it. Our servers also send PostHog a few events of their own, such as which cloud feature ran, and that switch does not cover those.
What we have and have not tested
What we did, 12 November 2026: we read Noted’s desktop code and its tests, the privacy page and the six vendor pages listed under Sources. We did not run a packet capture or the test suite for this post.
- Tested by hand by our founder on Mac and Windows, November 2026: dictated in Noted with Wi-Fi off, once signed in and the voice was downloaded. It shows dictation needs no connection; it is not a network check.
- One gate for cloud text. The code’s own comment says requests that carry your content to our AI
routes go through a single function,
post_llm, which refuses to build the request while Cloud is off. Before the app has read your settings, Cloud is off; if the settings file can’t be read, Cloud starts off. We did not trace every caller. - Cloud starts on. If no setting was ever saved, the code defaults to on. We haven’t checked how first-run screens present it, so look at Settings › General on a new install.
- Engine selection has no HTTP client. The file that picks and loads the speech engine makes no web requests; the voice model download lives in its own file. We did not audit the engine library itself.
- An idle connection you will see. With Cloud on, the app opens one bare connection to our API host when you are signed in (at launch, after sign-in and after the Mac wakes) to save time on your first request: a HEAD request with no token and no body. It is skipped while Cloud is off. If you run step 3 or 4 with Cloud on, expect it.
What nobody tested: the nettop check on any app, ours included; the Wi-Fi-off check on
any rival; what Apple
Dictation sends on any particular Mac; any rival’s behaviour beyond what its page says. If your measurement disagrees
with this post, trust your measurement and tell us.
How can I check where my dictation goes tonight?
About 20 minutes, and it works on any app. Expected results below are what each page says, not something we measured.
- Write down the claim. For each app, note what its settings say: Apple’s Keyboard settings text, a local or cloud model choice, a Cloud switch. Don’t test what you haven’t named.
- Wi-Fi off test. Install the app and finish any voice download. Turn Wi-Fi off and unplug Ethernet. Dictate into a plain text box. Text appearing shows speech to text doesn’t need a connection. It doesn’t prove nothing is queued to send later.
- Reconnect and watch. In Terminal, run
nettop -P(Apple’s built-in tool;-Pshows one line per process) and dictate a long sentence. Look for the app’s process and whether its bytes sent jump while you talk. - Take an idle baseline. Apps send small things (sign-in checks, analytics) when you aren’t talking. Audio is the difference: sustained upload only while you speak. Then look in the app for an analytics switch.
- One feature at a time. Dictate, then try clean-up, summaries and answers separately, so you know which one made the connection.
- Name the destination. A host that belongs to the app’s own service or its model provider is your cue to read that page’s privacy section. Then turn on FileVault if it isn’t on.
Our point of view
We think voice is the most personal thing a dictation app handles, so the line we drew is simple: audio stays on your Mac, with no cloud speech option at all. That is why Noted has no choice of voice model. It costs us flexibility, and we think the trade is right for a tool people use on client work and drafts.
We also think text features should sit behind one switch, not six, because one switch is easier to audit. The app routes content requests through one gate (above) and the same switch is in the menu-bar or tray icon, so you can turn it off before a sensitive session. And since Cloud can be on from the start, the honest advice is to look at it on day one.
Limitations, honestly
Where Noted is weaker. It has no choice of voice model and no cloud-free summary option, unlike Superwhisper, which documents local models for both a meeting’s transcript and its summary. Signing in and the first voice download need the internet; see Mac dictation with no internet for what works without it. Noted doesn’t encrypt its own files, so disk encryption is what protects your history if the Mac is lost; its privacy page names FileVault on Mac and BitLocker or Device encryption on Windows.
Where the built-in option wins. Apple Dictation needs no install and no new account, and when your Keyboard settings say it is processed on your Mac, it has no voice upload either. If that covers your need, use it.
Limits of the test. nettop and firewalls show connections and byte counts, not their
contents. “Not uploaded” covers audio; “private” also covers what is stored, who can read it and whether it trains a
model. For meeting notes in particular,
see meeting notes without a bot and
AI note takers and consent.
FAQ
Does Apple listen to my dictation?
Only if you chose to share audio recordings with Apple: its Dictation guide says that option lets Apple store audio of your Siri and Dictation use. You can turn that off in Analytics & Improvements.
Is Wispr Flow private?
Its pages say dictation is processed in the cloud, so audio leaves your Mac. The data-sharing toggle controls whether your content is used to improve models, and its AI providers may not train on your data.
Which Mac dictation never sends audio anywhere?
Apple Dictation when your Keyboard settings say it’s processed on your Mac, Superwhisper with a local voice model, OpenWhispr in local mode, and Noted’s speech to text. Check each app’s clean-up and summary steps separately.
Does turning Wi-Fi off prove an app is private?
No. It shows speech to text works without a connection. An app could still queue usage data or text to send when you reconnect, which is why the steps include a watch with the network on.
Does Noted store my dictation on a server?
No. Its privacy page says your dictation history stays on your computer. With Cloud on, Polish sends the text to be processed, not stored.
Can I stop Noted sending anything at all?
Turn Cloud off in Settings › General and no cloud feature sends anything, and turn off Share usage analytics in Settings › Privacy. Sign-in and the first voice download still use the internet, the app still asks PostHog which features are on for your account, and our servers may still log a few events of their own.
Sources
- Dictate messages and documents on Mac and If Dictation on Mac doesn’t work as expected (Apple Support)
- Manage data sharing, cloud storage and local history and Data usage (Wispr Flow)
- Sensitive data (Superwhisper)
- Where your data goes (OpenWhispr)
Checked 12 November 2026. Noted isn’t affiliated with Apple, Wispr, Superwhisper or OpenWhispr.
If you want to try one for yourself, hold one key and talk in any app: Noted Dictation.