Where Does Your Voice Go When You Dictate?

When you hold a shortcut and start talking, your words turn into text within a second. But behind that little bit of magic is a real question most people never ask: where does your voice actually go? The answer changes everything about your privacy, and it is not the same for every app.

The two paths your voice can take

Every dictation tool follows one of two routes. In the first, your microphone audio is packaged up and sent over the internet to a company's servers, where a large speech model transcribes it and returns the text. In the second, the speech model lives on your Mac, and the transcription happens on the same chip that runs your other apps. The words you see on screen can look identical either way. What differs is who else could, in theory, hear them.

This distinction is the heart of the question. If you care about privacy, the honest way to compare tools is not by how clean the output looks, but by whether the processing is genuinely on-device. A cloud tool can be excellent and still route every sentence you speak through a data center you do not control.

What happens with cloud dictation

Cloud dictation tools work because they borrow the power of large servers. When you speak, short chunks of audio are uploaded, transcribed remotely, and the text is streamed back to your cursor. This can produce very good results, and it is why some web-based tools feel so responsive. The trade-off is that a copy of your voice, at least momentarily, exists on infrastructure the vendor operates. Depending on the policy, that audio may be logged, retained, or used to improve models.

For a lot of everyday dictation, that may be fine. But it stops being fine the moment you dictate a client's medical history, an unpublished contract, a password read aloud, or anything covered by an NDA. In those cases the safest assumption is simple: if it leaves your Mac, treat it as if someone else could read it. That is exactly why we built BlaBlaType to never upload your voice in the first place.

What happens with on-device dictation

On-device dictation flips the model. The speech recognition engine, a local build of Whisper or Parakeet, ships inside the app and runs on your Mac's own processor. When you talk, the audio goes straight into that local model, gets converted to text, and is optionally cleaned up by on-device AI that removes filler words and fixes punctuation. At no point is a network request needed to turn your voice into words. You can confirm this yourself: switch off Wi-Fi, dictate a paragraph, and watch it still work.

Apple Silicon is a big reason this is now practical. The Neural Engine and unified memory let these models run quickly enough for real-time dictation, which used to require a server. The result is a setup where you get cloud-quality accuracy with none of the cloud's exposure. If you want the deeper mechanics, we cover whether Mac dictation is private in a separate guide.

How to tell where your voice really goes

You do not have to take any vendor's word for it. There are three concrete checks anyone can run:

• The airplane-mode test. Disconnect from the internet and try to dictate. If it keeps working, the transcription is local. If it fails or stalls, your audio was being uploaded.
• The privacy policy read. Search the policy for words like "transmit," "process on our servers," or "improve our models." Vague language about audio usually means the audio leaves your device.
• The network check. Tools like macOS Activity Monitor or a firewall app show whether the dictation process is sending data while you speak.

These checks matter because dictation is one of the most intimate things you can hand to software. You speak the way you think, and you often dictate faster than you would ever type. In fact, most people speak around three to four times faster than they type, which is precisely why voice input is so appealing, and why it is worth being deliberate about where those words end up.

Which route is right for you?

If you only ever dictate casual notes, either route is fine. But if privacy is a factor at all, on-device is the safer default because it removes the question entirely. There is nothing to audit, no policy to trust, and no server to worry about. Advanced keyboard-driven setups such as Talon Voice also run locally, though they are aimed at power users rather than everyday writing. For a friendlier on-device option that types everywhere and cleans up your speech, BlaBlaType is built for exactly this. It is also a strong pick among the best dictation apps for non-native speakers, since local models handle 90+ languages.

Voice input is only going to become more common as it gets faster and smarter. The tools worth adopting are the ones that give you the speed without asking you to give up control of your own words. Start by knowing where your voice goes, then pick accordingly. If you want to see plans first, they are on the pricing page.