When you dictate, your voice has to be turned into text somewhere. That “somewhere” — your own Mac, or a company’s servers — is the difference between on-device and cloud dictation, and it shapes everything from privacy to cost to how the tool feels.
01Where your voice actually goes
On-device: a speech model stored on your Mac processes the audio locally. Nothing is uploaded. The microphone stream becomes text on the same machine and is gone the moment it’s transcribed.
Cloud: the audio is streamed over the internet to a provider (Deepgram, OpenAI, ElevenLabs, and others), a large model transcribes it, and the text comes back. The connection is encrypted in transit, but the audio does leave your device and is processed on someone else’s infrastructure under their policies.
02Privacy: no contest, with nuance
For raw privacy, on-device wins decisively: data that never leaves your Mac can’t be logged, retained, subpoenaed, or breached on a server. There’s no “do they train on my audio?” question because there’s no upload.
Cloud isn’t automatically reckless — reputable providers offer encryption, retention controls, and zero-retention modes — but you’re trusting a third party and their configuration. If you handle confidential, medical, legal, or NDA-bound material, on-device removes an entire category of risk rather than mitigating it.
03Accuracy: closer than you’d think
Cloud models are large and frequently updated, so on the hardest inputs — heavy accents, noisy rooms, specialized jargon — they often still lead. But the gap has narrowed sharply. Apple’s on-device models, including the upgraded macOS 26 Speech, handle everyday dictation cleanly, and for clear speech in a quiet room you may not notice a difference at all.
A useful rule: judge accuracy on your voice and vocabulary, not benchmarks. Punctuation, capitalization, and proper-noun handling affect editing time more than a fraction of a percent of word-error rate.
04Latency, cost, and offline use
- Latency: on-device avoids a network round trip, so it can feel instant; cloud depends on your connection.
- Cost: on-device is free to run. Cloud is usually pay-per-minute (often through your own API key), which is fine occasionally and adds up if you dictate all day.
- Offline: on-device works on a plane, a train, or a bad hotel Wi-Fi. Cloud simply doesn’t.
05You don’t have to choose forever
The smartest setup uses on-device as the private, zero-cost default and keeps a cloud engine in reserve for the few recordings that genuinely need a bigger model. Better still is choosing per language: let an on-device engine handle the language it’s great at, and route a trickier language to the cloud provider that does it best.
Have both, on your terms
VTT defaults to on-device and private. Add Deepgram, OpenAI, or ElevenLabs with your own key only when you want it — and pick the engine per language.
Download VTT06A simple decision guide
- Sensitive content, or you just prefer not to upload your voice → on-device.
- Maximum accuracy on a hard recording, and the content isn’t sensitive → cloud, ideally with a zero-retention setting and your own key.
- Both, depending on the moment → a tool that lets you switch (or auto-route by language) without re-recording.