Is voice dictation private? What happens to your words

Updated: November 7, 2026

Every dictation tool has to turn sound into text somewhere. That somewhere is the whole privacy question: your own processor, a server owned by the vendor, or a mix where the audio stays home but the text travels. This guide explains the three models in plain terms, gives you the questions that make a vendor answer precisely, shows how to check a tool yourself in five minutes, and ends with a checklist you can run before dictating anything you would not post publicly.

Three places your voice can be processed

Marketing pages rarely say it directly, so it helps to know the vocabulary. Local (also called on-device or offline) means the speech model runs on your computer and the audio never needs a network connection. Online (or cloud) means the recording is sent to a server, converted there, and the text is sent back. Hybrid means one step happens locally and another remotely, most often local transcription followed by an optional online clean-up of the text.

None of the three is automatically safe or unsafe. A local app on a shared, unencrypted laptop can leak more than a well-run service. The point is to know which model you are using, not to assume.

What the built-in options tell you

On a Mac, Apple says the Keyboard settings show whether your dictation is processed on the device or sent to its servers, and that sharing audio recordings with Apple is a separate, optional choice. Open System Settings > Keyboard > Dictation and read the line under the switch: it depends on your language and Mac, so check it on your own machine. The Mac dictation guide walks through that pane.

On Windows, Microsoft states that the dictation built into Windows relies on online speech recognition, so it needs an internet connection. In other words, it is an online model by design. If that is a problem for your work, it is a reason to look at a local tool rather than a setting to toggle; the Windows dictation guide covers the setup.

Questions to ask any dictation vendor

Send these by email or look for them in the privacy policy. A vendor who answers all of them in writing is easier to trust than one who answers with a badge.

• Where is speech converted to text? On my device, on your servers, or both? If servers, in which country?
• Is audio stored after conversion? If yes, for how long, and can I delete it?
• Is the resulting text stored? Many policies cover audio and say nothing about text.
• Is any of my data used to train models? Is that opt-in, opt-out, or not done at all?
• What else is sent? App names, window titles, surrounding text for context, clipboard contents?
• Which features need the internet? Rewriting, translation and formatting often do, even in apps that transcribe locally.
• Is there a fully offline mode, and what stops working in it?

Check a tool yourself in five minutes

Answers in a policy are one thing; behaviour is another. A quick practical test tells you whether a tool depends on the network for transcription.

• Open a plain text editor and dictate a short test sentence: "The blue folder is on the second shelf." Confirm it works normally.
• Turn off Wi-Fi and unplug Ethernet. Dictate the same sentence again.
• If text still appears, transcription is local. If you get an error or nothing happens, the tool uses a server.
• With the network still off, try any extra features (rewrite, summarise, translate). Whatever fails offline is sending something online when you are connected.
• Turn the network back on and note whether the app asks to sync, upload or log in again.

This test shows where transcription happens, not what a vendor does with data it does receive. For that you still need the written answers above. Offline dictation on Mac adds a second test that watches network traffic.

A checklist before you dictate sensitive text

• Know which of the three models your current mode uses, and whether a setting can silently change it.
• Read your employer's, school's or client's rules about third-party software. Those rules win over any vendor promise.
• If you use an online service, prefer one that states in writing that it does not store audio or text, and keep its answer on file.
• Turn off optional online rewriting for confidential documents.
• Watch where the cursor is. Dictation types into whatever field has focus, including a chat window you forgot was open.
• Lock the computer when you step away. A local model is only as private as the device it runs on.
• Use full-disk encryption (FileVault on Mac, BitLocker or device encryption on Windows) so a lost laptop does not expose your drafts.

What Voice Monkeys does with your voice

Voice Monkeys is a local tool on every platform it supports: Macs with Apple Silicon (M1 or newer, macOS 14+), Macs with Intel (macOS 11+, 8 GB of memory recommended), Windows 10 or 11 64-bit and 64-bit Linux. Speech is recognized on the computer itself. The internet is needed once to download the speech model (about 600 MB on Apple Silicon, about 640 MB elsewhere), then only for the licence check and updates. No audio or text is sent.

You can confirm that with the Wi-Fi-off test above. The price is $25 once for a lifetime licence on one computer, with a 7-day free trial without a card, so you can download Voice Monkeys and run the test before paying.

Which model fits which situation

Profession-specific advice is in the guides for doctors and lawyers, and how to choose dictation software covers the other criteria.

Questions and answers

Sources

• Apple: Use Dictation on Mac (checked November 7, 2026)
• Microsoft: Speech, voice activation, inking, typing, and privacy (checked November 7, 2026)

More guides

• Dictation for doctors: a workflow for clinical notes
• Dictation for lawyers: drafting memos and emails by voice
• Offline dictation on a Mac
• How to choose dictation software: the criteria that matter