Most built-in and popular dictation tools use cloud speech recognition: your recorded voice is uploaded, transcribed on the vendor's servers, and the text comes back. Some tools, including Apple Dictation on supported Macs and open-source Whisper, run the model on your device instead. Neither approach is a secret, but vendors describe it in different places, so this guide puts the answers side by side.
We make AirTypes, which does both, and says which is which. Plain is free and offline: Whisper runs on your CPU and 0 bytes of audio are uploaded. Polished cleans your words up in the cloud at 120 minutes for every dollar, and it sends the recording to AirTypes to do it. Here is what each typed from the same sentence:
"ok so um the meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again"
The meeting is moving to tomorrow at ten, client is not available, and uh send him the invoice again.
Polished typed: "The meeting is moving to tomorrow at 10:00. The client is not available. Send him the invoice again." Cleaner text, but the recording left the machine. That trade is the whole subject of this article.
Cloud vs on-device speech recognition
Every dictation tool does the same job: it records a few seconds of your voice and turns that sound into text. The privacy difference comes down to one step, which is where the speech model runs.
Cloud speech recognition
The app streams or uploads your audio to the vendor's servers. A large model transcribes it there and sends the text back. This is why cloud tools stop working when you lose your connection. Once audio reaches a server, the vendor's privacy policy decides what happens to it: whether it is stored, for how long, and whether it can be used to improve their models. Many vendors offer settings for this, so the answer can differ by account.
On-device speech recognition
The speech model is a file on your computer. Your audio goes from the microphone to that model and becomes text without a network round trip. Modern open models such as OpenAI's Whisper made this practical on ordinary laptops. If you want the technical detail, our guide to offline speech recognition walks through how on-device recognition works and how the engines compare.
What "on-device" guarantees, and what it does not
On-device recognition means the audio does not need to leave your machine to be transcribed. That is a strong property, but it is not the whole picture:
- The app can still send other data. Account details, crash reports, analytics or usage statistics may go to the vendor. Read the privacy policy to see what.
- The destination app has its own rules. If you dictate into a cloud document or a chat app, the text you produce is stored by that service, exactly as if you had typed it.
- Optional features can use the network. AI rewriting, sync or translation features may send audio or text somewhere, even when basic transcription is local.
So "on-device" answers the audio question. The privacy policy answers everything else.
Where common dictation tools process audio
The table below summarises what each vendor's own documentation and product pages describe. Vendors change their products, so treat it as a starting point and check the current privacy page before you rely on it.
| Tool | Where speech is processed | Works offline? |
|---|---|---|
| Apple Dictation (macOS) | Apple's documentation states that on Macs with Apple silicon, supported languages are processed on the device. On other setups, audio may be sent to Apple. | On supported Macs and languages |
| Windows voice typing | Microsoft's documentation describes it as using Microsoft's online speech recognition service. | No, it needs an internet connection |
| Google Docs voice typing | Runs in the Chrome browser and uses Google's cloud speech recognition. | No |
| Wispr Flow | Transcribes in the cloud only. It offers a Privacy Mode setting. | No |
| Otter.ai | Cloud transcription service built for meetings; recordings are uploaded and processed remotely. | No |
| Superwhisper | Local Whisper models on your device; its Pro cloud models send audio off the device. | With local models |
| Whisper (open source, run locally) | On your device. The model runs on your own CPU or GPU. | Yes |
| AirTypes Plain | On your device. Whisper runs on your CPU; 0 bytes of audio uploaded. | Yes |
| AirTypes Polished | Polished sends the recording to AirTypes to transcribe and clean it up. | No, it needs a connection |
A few notes on reading this table. Apple's split means two Mac users can get different answers to "is Apple dictation private": check which chip your Mac has and whether your language is supported for on-device dictation. Windows includes other speech features that work differently from voice typing, so check the documentation for the specific feature you use. And "Google voice typing" can also mean Google's mobile keyboard, which is a separate product with its own settings on your phone.
If you want a deeper side-by-side, we have pages on Apple Dictation alternatives, Windows voice typing alternatives and a direct AirTypes vs Wispr Flow comparison.
Dictation that can stay on your machine
AirTypes Plain runs Whisper on your CPU and types at your cursor in any app. Free on macOS and Linux, no card, no time limit.
Download AirTypesFive questions to ask any dictation tool
You do not need to trust a marketing page, including ours. These five questions cover what matters, and a vendor's privacy policy or documentation should answer each one.
-
1
Where is my audio processed?
On my device, or on the vendor's servers? If it is "both depending on settings", find out which setting you have.
-
2
Is my audio or transcript stored?
If it is, where, for how long, and can I delete it?
-
3
Is it used to train or improve models?
Look for an opt-in or opt-out setting, and whether people may review samples.
-
4
Does it work offline?
A tool that transcribes without a connection cannot be uploading audio to transcribe it.
-
5
What else is sent?
Account details, analytics, crash reports, and anything optional AI or sync features send.
How AirTypes handles your voice
Here are the same answers for AirTypes, stated as plainly as we can. AirTypes has two modes, and you switch between them on the app's Home screen.
Plain: transcription runs on your CPU
You hold a hotkey (Ctrl + Shift + Space by default), speak, and let go. Whisper transcribes the recording on your own processor, light cleanup on the machine removes a leading um or uh, capitalizes the first letter and adds a period, and the text is typed at your cursor. 0 bytes of audio are uploaded, and it works in airplane mode. You pick one of two Whisper model tiers: Swift (about 78 MB), the fastest, or Crisp (about 148 MB), the one we recommend. Plain is English only, and free with no time limit.
Polished: the recording goes to AirTypes
Polished sends the recording to AirTypes to transcribe and clean it up. A large speech model transcribes it on our servers, then an AI removes filler words and false starts and fixes grammar and punctuation, keeping your meaning, numbers and names. It works in the language you speak, and it needs an internet connection. It is prepaid ($5 buys 600 minutes of speech) with no subscription. If you never switch to Polished, no recording is ever sent. What happens to a Polished recording is set out in our privacy policy. See how Polished works.
What AirTypes receives
You create a free account at first launch. AirTypes always receives your account email and usage counts. In Plain, nothing else. In Polished, it also receives the recording, as described above.
Honest limits
AirTypes is a dictation tool, not a meeting recorder. Plain is English only; other languages need Polished, which is cloud based. The desktop app is live on macOS and Linux, and the Windows app is in development (you can join the Windows waitlist). On Windows today you can Dictate in your browser (new in November 2026), but be clear about what that is: Dictate is Polished only, so every recording is sent to AirTypes. Only Plain in the desktop app keeps your voice on your computer.
On-device or cloud: how to decide per task
Privacy is not all or nothing. Most people dictate a mix of things: a note to themselves, a reply to a client, a prompt for a chatbot, a paragraph of a report. The useful question is not "which tool is private?" but "where should this recording go?"
| On-device (AirTypes Plain, local Whisper) | Cloud (AirTypes Polished, Wispr Flow, built-in cloud dictation) | |
|---|---|---|
| Where the audio goes | Nowhere; it stays on your computer | To the vendor's servers |
| Works offline | Yes | No |
| Cleanup | Light: words as you said them | Can rewrite fillers, false starts and grammar |
| Languages | Depends on the model (Plain: English) | Usually many |
| Load on your laptop | Runs the model locally | Light; the server does the work |
| What you rely on | Your own machine | The vendor's privacy policy and settings |
Stay on-device when
- The words belong to someone else: client matters, patient details, unreleased work, sources.
- A contract, policy or regulation says where data may be processed.
- You are offline, or on a network you do not trust.
- You simply would rather not have the recording leave your machine.
Cloud is reasonable when
- You want finished text you can send without editing, and the content is not sensitive.
- You dictate in a language your on-device model does not cover.
- Your laptop is too slow to run a good local model.
For example, a quick update to a teammate is a fine fit for cloud cleanup. Here is a real Polished result:
"so um I think we should uh probably move the deploy to friday because like the migration isn't, the migration isn't tested yet and uh QA needs two more days"
I think we should move the deploy to Friday because the migration isn't tested yet and QA needs two more days.
Notes on a client's legal matter are not. With most tools you make this choice once, when you pick the app. With AirTypes you make it each time: Plain for the sensitive recording, Polished for the one you want cleaned up, switched on the Home screen.
Who should care most
For a grocery list, cloud dictation is usually fine. The calculation changes when what you say belongs to someone else, or when a rule says where it may go.
- Lawyers dictating notes, correspondence or case details covered by confidentiality.
- Clinicians writing notes that include patient information.
- Developers describing proprietary code, architecture or security issues.
- Journalists working with sources and unpublished material.
- Anyone with GDPR, HIPAA or contractual obligations about where personal or client data is processed.
On-device transcription removes one data flow from the list you have to account for, which can make a review simpler. It does not make any tool compliant by itself, and this article is not legal advice: check with your compliance team about your obligations. For an organisation-level view, read our guide to enterprise offline voice recognition.
A two-minute privacy test
You can check the core claim yourself, for any tool, without reading a single policy.
-
1
Turn off Wi-Fi and unplug Ethernet
Or switch on airplane mode, so the computer has no network at all.
-
2
Open a local text editor
Use something that does not need the internet itself, such as TextEdit, gedit or a code editor.
-
3
Dictate a sentence
If the text appears, recognition is running on your device. If the tool shows an error or waits for a connection, it relies on a server.
This test tells you where transcription happens. It does not tell you what an app sends once you reconnect, which is why the privacy policy still matters. If you try AirTypes this way, use Plain mode; Polished needs a connection by design. The AirTypes setup guide covers hotkeys and model tiers.
FAQ
Is voice dictation recorded?
Every dictation tool captures your microphone audio, because it has to hear you to transcribe you. The real question is where that audio goes next. A cloud tool sends it to a server for transcription, and the vendor's privacy policy decides whether it is kept. An on-device tool transcribes it on your computer and has no need to upload it.
Does Apple dictation send audio to Apple?
It depends on your setup. Apple's documentation states that on Macs with Apple silicon, dictation in supported languages is processed on the device. On other Macs or languages, dictation may be sent to Apple for processing. Check Apple's current dictation and privacy pages for your exact Mac and language.
What is the most private dictation app?
The most private option is one that runs speech recognition on your own device, so audio never has to leave it. Open-source Whisper run locally is one route. AirTypes Plain is a desktop mode built on the same approach: Whisper runs on your CPU and 0 bytes of audio are uploaded. Whatever you pick, read its privacy policy to see what else it sends.
Does AirTypes see what I dictate?
It depends on the mode you choose. In Plain, no: transcription runs on your computer, and AirTypes receives only your account email and usage counts. In Polished, yes: Polished sends the recording to AirTypes to transcribe and clean it up. You switch modes on the Home screen, so you decide per recording.
Get AirTypes, free
Plain keeps your voice on your own computer, free and offline. Polished is optional, for the recordings you want cleaned up in your language. The desktop app runs on macOS and Linux now; the Windows app is in development.
Download for macOS and Linux