On-device vs cloud transcription: what actually leaves your phone

Every speech-to-text app on iPhone that does not ship its own model is calling the same thing under the hood: Apple’s Speech framework. That framework has two modes, and the difference between them is the whole privacy story.

The two modes, concretely

A speech recognition request on iOS carries a flag that asks for on-device recognition. When it is on, the audio is decoded by a model that already lives on your phone, downloaded as part of the language assets. Nothing goes out over the network — the request completes in Airplane Mode.

When the flag is off, the audio is streamed to Apple’s speech recognition service, transcribed there, and the text comes back. This is the older path and, for many languages, still the more accurate one, because the server model is far larger than anything that fits in a phone’s memory budget.

The trap is that the flag is a request, not a guarantee. If the requested language has no on-device model installed, an on-device request simply fails or returns nothing. Apps then have a choice: report failure to you, or silently retry against the server.

What determines which one runs

Language coverage

Apple ships on-device recognition for a subset of the languages it supports overall. Widely used ones — English, Spanish, French, German, Mandarin, Japanese — have local models on current hardware. Smaller languages and some regional variants are server-only. This list changes between iOS releases and is not something an app controls.

Whether the language asset is downloaded

Supported is not the same as installed. The on-device model for a language arrives when the system decides to fetch it, which usually means after you have added that language to your keyboard or dictation settings and the phone has had time on Wi-Fi. A freshly reset phone can be "supported" and still have nothing local to run.

Audio quality and length

Local models are smaller and less tolerant. Quiet, distant, heavily accented or noisy audio that a server model handles will sometimes come back empty from the local one. Long files get split into chunks by most implementations, and an individual chunk can fail while the rest succeed.

How to tell which path actually ran

There is no badge in the UI of most apps. The reliable test takes thirty seconds and does not require trusting anyone’s marketing:

• Turn on Airplane Mode and leave Wi-Fi off.
• Record a short clip and run the transcription.
• If you get text back, that text was produced on your device. There was no network to use.
• If you get nothing, or an error, that language or that audio was going to the server.

Repeat it per language. The result for English tells you nothing about the result for Ukrainian.

What is actually transmitted on the server path

When the server path runs, the audio is what gets sent — not a summary, not just the text. Apple documents this as part of its Speech framework and its privacy disclosures, and it is the same service that has always powered dictation on the keyboard. It is a first-party Apple path, not a third-party one, but "first-party" is not the same as "local", and the two get conflated constantly in app descriptions.

Be sceptical of any app that says "fully on-device" without qualifying it by language. If it uses the Speech framework and it supports forty languages, it is not fully on-device on all forty. Either it fails on the rest, or it falls back.

How Voice Studio handles it

Voice Studio uses the Speech framework and always attempts the on-device path first. If that pass returns nothing or errors out, it retries the same audio without the on-device flag, which means the audio goes to Apple’s speech service when you are online. That fallback exists because the alternative is handing you an empty transcript, and an empty transcript is what people report as "the app is broken".

We would rather say that plainly than claim a guarantee we cannot keep. If you need certainty for a particular recording, transcribe it in Airplane Mode: the fallback has no network to reach, so either the local model handles it or you get nothing — and either way nothing was sent. Beyond that path, Voice Studio has no account, no server of its own, and no analytics SDK; recordings and transcripts are stored in the app’s own storage on the device.

Common questions

Does on-device transcription work in Airplane Mode?

Yes, for languages that have an on-device model installed on your phone. That is exactly why Airplane Mode is a useful test: if text comes back with no network available, it was produced locally.

Is Apple’s server transcription the same as sending my audio to a third party?

It is a first-party Apple service, the same one behind keyboard dictation. It is not a third-party company, but the audio does leave your device, so it is not local either.

Can an app force on-device only and never fall back?

Technically yes — it can request on-device recognition and simply report failure instead of retrying. The cost is empty transcripts for any language or audio the local model cannot handle.

Does Voice Studio ever send my recordings to its own server?

No. Voice Studio has no account and no server of ours. The only path on which audio leaves the device is Apple’s speech service, as the fallback described above.