Does on-device transcription work in every language on iPhone?

It is easy to assume "on-device transcription" is a feature an app either has or does not have, the same way it either has dark mode or does not. It is not that simple. On-device recognition is built language by language, and iOS only ships a local model for some of them. An app can support the on-device path perfectly and still send audio to a server, purely because of which language was spoken.

Why every language does not get a local model

A speech model that runs entirely on a phone has to fit in a few hundred megabytes and still work with no server to lean on. Building one that is accurate takes a large amount of training data and engineering effort, per language, per region. Apple has done that work for the languages with the most speakers and the most demand, and has not — yet, or possibly ever — done it for every language and regional variant it supports through its speech framework.

That is a decision made at the iOS level, not the app level. No app can add a local model for a language that Apple has not shipped one for, and no app can be blamed for a gap it has no way to close.

Two different reasons a language ends up on the server

It never had a local model to begin with

For these languages, an on-device request is not a slower or less accurate path — it is not a path at all. It fails immediately, every time, regardless of how quiet the room is or how clearly you spoke. The only way to get a transcript in that language is the server path, when you have a connection.

It has a local model, but this particular recording defeated it

This is a separate, more common situation. The language does have an on-device model, but local models are smaller and less forgiving than server models — quiet audio, heavy background noise, or a strong accent can make the local pass come back empty even though the language itself is supported offline. That is a per-recording failure, not a per-language one, and a cleaner recording of the same language might transcribe locally without issue.

Widely spoken languages first

In practice, the languages with the largest number of speakers — English, Spanish, French, German, Mandarin and Japanese among them — have on-device models on current iPhones. Many smaller languages and some regional variants do not, and are server-only until Apple adds them. That list is not published anywhere permanent and it changes between iOS releases, so a language that is server-only today is not guaranteed to stay that way.

Why this matters beyond speed or accuracy

The usual framing for on-device versus server transcription is privacy: does the audio leave the phone. For a language with no local model, that question has a fixed answer — yes, whenever a transcript is produced at all, because there is no other way to produce one. Putting the phone in Airplane Mode does not change that; it just means the request has nowhere to go, so you get nothing back instead of a transcript. That is a genuinely different situation from a widely-spoken language that usually stays local and occasionally falls back — for a server-only language, "occasionally" is "always".

How Voice Studio handles this

Voice Studio always attempts the on-device path first, for every one of its 38 transcription languages, and only retries over the network if that pass comes back empty or throws an error. For a language with no local model, that first attempt fails immediately and the retry is what actually produces the transcript when you are online — the app is not choosing to skip the local path, iOS simply has nothing local to offer it. There is no way to force a transcript for one of those languages without a connection, because there is no local model to force. Beyond that fallback, Voice Studio has no account and no server of its own; recordings and transcripts stay in the app’s own storage on the device either way.

Common questions

Which languages have on-device transcription on iPhone?

There is no single published list, and it is Apple’s decision rather than any app’s. Widely spoken languages tend to have it; smaller languages and some regional variants often do not yet. It also changes between iOS releases.

What happens if I use Airplane Mode to transcribe a server-only language?

You get nothing back — not a slower transcript, an empty one. With no local model and no network, there is no path left for the request to take.

Does supporting 38 transcription languages mean all 38 run on-device?

No. That count is how many languages the app can request transcription in. How many of those iOS can actually run locally is decided by Apple and is smaller than the full list.

Can I find out in advance whether my language needs a connection?

Yes: in Airplane Mode, open Notes, tap the keyboard’s microphone, and dictate a sentence in that language. If text appears, that language has a local model on your phone. If nothing happens, it does not.