Does dictation send your audio to a server?
For most dictation apps, yes. The audio goes to a machine you do not own, is transcribed there, and is kept for some period you are not told. The short version of this post: it depends on whether the app has a speech model on your disk.
How to tell, in about thirty seconds
Three checks, in order of reliability.
1. Turn off the network. Disconnect Wi-Fi and Ethernet, then dictate. If text still appears, the engine is local. If the app stops responding, or errors, or shows "connecting", it was uploading. This is the only test that cannot be faked by marketing copy.
2. Look for a model download. A local engine needs weights on disk — hundreds of megabytes, downloaded once. An app that is a 20 MB download and works instantly on first launch is almost certainly calling an API.
3. Read the privacy page for the word "transcribe". "We transcribe your audio" is the giveaway; the question is where. "Processed on your device" is a claim, and check 1 is how you verify it.
What cloud dictation does with the audio
Typically: the file is uploaded, transcribed, and the transcript is returned. Then one of three things happens to the audio, and vendors rarely say which.
- Deleted immediately. The most common claim, and the hardest to verify.
- Kept for a window. Commonly 7 to 30 days, for abuse review and debugging.
- Kept to improve the model. Often opt-in, sometimes the default for a free tier.
All three mean the same thing in practice: a copy exists somewhere you cannot see, on a machine you cannot examine, governed by a policy that can change.
Why it matters more for dictation than for search
Dictation is not a search query. People dictate messages to their family, medical notes, contract language, code comments, and complaints about their employer. It is the most candid text they produce all day, and it arrives at the server unedited, with their voice on it.
A voice recording carries more than the words: accent, age range, gender, health, and background noise that can identify a room. That is why "we delete it after processing" is a weaker promise than "it never left".
What a local engine changes
The audio is decoded on the same chip that is running your display. Nothing is uploaded because there is no upload path in the code. The trade is real and worth stating plainly: a cloud service with a large proprietary model will usually beat a model that fits in your laptop's memory. You are choosing privacy over a couple of points of word error rate.
For most dictation — messages, notes, prompts, commit messages — the local model is good enough, and the gap closes with a custom vocabulary list for the names and jargon you actually use.
Frequently asked
Is Apple's built-in dictation local? Partly. On recent macOS versions some of it runs on-device, but it is not a promise you can audit, and it is not configurable.
Can a local app still leak? Only if you enable something that sends text: an optional cloud rewrite, or a feedback report. In Inmost both are explicit and off by default, and the app never sends audio.
Does local use more battery? Yes, briefly, while the model decodes. It is a few seconds of work per dictation on Apple Silicon.
Inmost is a local-first dictation app for macOS with no account and no upload path for audio. See how it compares.