Where does iPhone Voice Memos transcription happen?

Apple does not say. Its Voice Memos transcription page lists the requirements — iPhone 12 or later, and ten languages — but never states whether the work happens on the phone or on Apple's servers. Apple is explicit about the keyboard instead: dictation requests are processed on your device in many languages, with no internet connection required. If it matters for a particular recording, airplane mode will settle it in a minute.

Do you have to download a language pack to transcribe offline?

Usually, yes — offline language support means files sitting on the phone, so each language is a download rather than a setting you flip. Apple's Translate app works exactly this way: to translate without a connection you download both the input and output languages first, and Apple notes that offline translations may not be as accurate as online ones. Check your language is in the downloadable set before you rely on it somewhere with no signal.

Why does my voice app transcribe in the wrong language?

Because most transcription follows a language setting rather than listening for which language is being spoken, and on an iPhone that setting is usually the phone's own language. Record German on a phone set to English and the transcript comes back as nonsense that looks like English. Check whether the app has its own transcription language picker, separate from the system language, before assuming the audio was the problem.

Do AI note features need an internet connection?

Almost always, yes. Recording, transcription and ordinary word search can run on a modern phone, but the step that reads across your notes and writes an answer in plain language generally runs on a server, which means some of your text leaves the device. That is a separate question from where transcription happens, and worth checking separately.

How do you keep notes private while using AI?

Ask any app two questions, in this order: what leaves the device, and what whoever receives it keeps. Anything that runs on the phone answers the first by construction — nothing was sent, so there is nothing to retain. For the features that do use a server, look for a plain statement of what is sent and how long it is held; "encrypted" answers neither question, because it describes the journey rather than the destination. If an app will not answer both in plain words, treat the second answer as unknown rather than assuming it is nothing.

What is on-device voice transcription

On-device voice transcription means the phone turns your speech into text itself, using a model stored on the device, so the recording is never uploaded and the app keeps working in airplane mode, on a plane or in a dead zone. Cloud transcription sends the audio to a server instead, which usually handles difficult recordings and rare languages better, but needs a connection and means a copy of your voice leaves the phone. Most voice-note apps are a mix of the two — capture and transcription on the device, and anything that answers questions across your notes on a server. The quickest way to find out which kind you are using is to turn airplane mode on and record something.

What does on-device transcription mean?

It means the phone turns your speech into text itself, using a model that is already stored on the device, and the audio is never uploaded to do it.

Cloud transcription does the same job somewhere else. The app sends the recording — or streams it as you talk — to a server, a much larger model transcribes it there, and the text comes back down the wire.

The two are indistinguishable while you are standing still with five bars of signal. They stop being indistinguishable the moment the signal drops, or the moment you record something you would not email to a stranger.

Apple says plainly which one the iPhone keyboard uses. From the iPhone User Guide:

Dictation requests are processed on your device in many languages—no internet connection is required.

“Many languages” is the caveat that explains why both kinds still exist, and the same page notes the exception that proves the rule: “When dictating in a search box, dictated text may be sent to the search provider in order to process the search.” One feature, two paths, depending on what you are dictating into.

Does voice-to-text work in airplane mode?

On-device transcription does. Cloud transcription does not — it fails, or it quietly queues the recording until you have signal again.

That difference is the honest test, and it takes a minute:

Turn on airplane mode, and check Wi-Fi is off as well — airplane mode switches it off, but phones remember it and turn it back on.
Record thirty to sixty seconds of ordinary speech in the app you are assessing.
Watch what happens when you stop. Text within a second or two, with no spinner, means the work happened on the phone.
Read the failure carefully if it fails. “Will transcribe when you’re back online”, a permanent spinner, or an error mentioning the network all mean the audio was on its way to a server.
Then try the rest of the app offline — search your existing notes, and try whatever feature answers questions for you. This is where most apps split in half.

Step five matters more than the other four. The common case is not a purely offline app or a purely online one. It is an app that captures, transcribes and searches on the phone, and reaches for the internet only when you ask it something in plain language.

How can you tell which kind of transcription an app uses?

Airplane mode is the reliable answer, because it tests behaviour rather than marketing copy. But there are two other places to look.

The phone will sometimes tell you. Apple’s Siri, Dictation & Privacy notice describes exactly this:

When you use Dictation, your device will indicate in Keyboard Settings if your audio and transcripts are processed on your device and not sent to Apple servers. Otherwise, the things you dictate are sent to and processed on the server, but will not be stored unless you opt in to Improve Siri and Dictation.

Read that twice, because it says something people rarely expect: the same feature on the same phone can run either way, and Apple’s answer to “which one am I getting?” is to go and look in Keyboard Settings. On-device is not a badge an app wears permanently. It is a description of one path through one feature.

The requirements list is a fingerprint. Apple’s page for Voice Memos transcription sets out the conditions:

Audio transcription is available on iPhone 12 or later in English (all variants), Spanish, Portuguese, Italian, French, German, Japanese, Korean, Simplified Chinese, and Traditional Chinese.

A hardware floor and a ten-language list is the shape work done on the phone usually takes, because a server does not care which iPhone you own. Treat it as a hint rather than a verdict, though: unlike its keyboard dictation page, Apple does not actually say on this one where the transcription runs.

One caveat worth holding on to: airplane mode tells you where the transcription happened, not where the text ends up afterwards. An app can transcribe locally and still sync every word to an account the moment you land. Those are two separate questions, and you have to ask both.

Why does on-device transcription support fewer languages?

Because every language it handles is a file that has to fit on your phone, be kept up to date, and run fast enough to feel instant.

Servers have no such limit, which is why cloud services advertise language counts in the dozens and on-device features tend to list ten or so. Apple’s Translate app makes the trade explicit — you download the languages you want in Settings, and Apple’s own note there is refreshingly blunt: “Offline translations may not be as accurate as online translations.”

The practical consequence is the complaint you see over and over in Apple’s own forums: transcription follows a setting, not the sound. It does not listen and work out which language you chose to think in.

A September 2024 thread about Voice Memos, marked “Me too” more than 500 times, was opened by someone who recorded in Chinese, got garbled English back, and asked where the button to select the transcription language is. In April 2025 another user, recording German meetings, put the same problem in one line: “the transcription does not work because the phone language is English.”

So if you think in one language and set your phone in another, the question to ask an app is not “does it support my language” but “does it let me choose my language, separately from the system.” Those are different features, and only one of them helps you.

Is on-device transcription less accurate?

Sometimes, and the honest answer is that it depends entirely on the audio.

Server models are bigger, run on hardware with no battery to protect, and can be updated whenever the vendor likes. That advantage shows up most on hard audio: strong accents, background noise, two people talking over each other, technical vocabulary, a phone in your pocket. For one person speaking clearly into their own phone, the gap is frequently too small to care about.

Be suspicious of any accuracy percentage quoted at you, including a flattering one. Published figures come from benchmark recordings that sound nothing like your kitchen, your car or your accent, and none of them predict what you will get. The only measurement that means anything is yours:

Record your hardest case first — the noisy one, the fast one, the one full of names.
Record a normal one — the sort of thought you would actually capture walking to the shops.
Read both transcripts, not the summaries. Summaries paper over transcription errors beautifully, which is exactly the problem.

If the hard case is unusable and the normal one is fine, you have learnt something more useful than any comparison table: use the app for thinking out loud, and use something else for the difficult recordings.

What still needs the internet?

The asking. Almost everything else in a personal voice-note app can run on a modern phone; the step that reads across everything you have saved and writes an answer in plain language generally does not.

What you are doing	On-device app	Cloud app
Recording	Works anywhere	Works anywhere — the audio waits on the phone
Turning speech into text	Works in airplane mode	Needs a connection, or queues until it has one
Finding a note by word	Works offline	Depends where the notes are stored
Asking a question in plain language	Usually needs a connection	Needs a connection
Where your voice goes	Stays on the phone	A copy is uploaded and transcribed on a server

This is not a scandal, and cloud transcription is not a trap — it is how most services work, and for most recordings most people are entirely comfortable with it. What matters is that the answer is knowable. An app that is straight with you will say which text leaves the device, at what moment, and whether anyone keeps it. An app that only says “secure” and “encrypted” has answered a different question from the one you asked.

Which one should you choose?

Choose on-device when the recording itself is the sensitive part, or when the signal is not guaranteed:

Health, money, clients, legal matters, anything about other people — the strongest privacy guarantee is an audio file that never left the phone.
Planes, trains, basements, lifts, rural roads, abroad without roaming — the places thoughts arrive are frequently the places with no bars.
You dislike accounts. On-device features tend not to need one, because there is no server to authenticate you against.

Choose cloud when the audio is the hard part and the content is not sensitive:

Meetings with several speakers, where you want speaker labels and long recordings, which is a genuinely different job. Where a meeting archive stops working as a personal memory is worth reading before you pick a side.
A language the on-device list does not cover. No amount of privacy helps if the transcript is nonsense.
Recordings that must be right — interviews, anything quoted, anything with a record-keeping purpose.

Most people do not have to pick once and for all. Capture the everyday thinking on-device, and keep a cloud tool for the handful of recordings that genuinely need it.

When this does not fit

You have an older iPhone. Apple’s Voice Memos transcription starts at iPhone 12, and third-party on-device features have floors of their own. Cloud services do not.
You record for the record. Journalism, research, compliance and legal work need the original audio and the most accurate transcript available, and consent rules matter more than either.
You never leave signal and do not mind the upload. Then this whole question is theoretical for you, and optimising for it will cost you features for no gain.
You need it on a laptop too. On-device means on that device, which is a real limitation and not a small one.

And whichever you choose, re-test it after a big update. Vendors move features between the phone and their servers more often than they announce it, and airplane mode will tell you in under a minute — which is the same return-trip discipline that keeps a pile of voice memos from rebuilding, applied to the tool instead of the notes.

Fly Notes is a voice note-taking app for iPhone, and it tries to answer the question above in the same terms rather than deflect it with the word “secure”: capture, organizing and search run on your iPhone. Ask is the part that needs a connection — it sends your text to our AI providers (via OpenRouter) to answer it, zero-data-retention providers only. We tell you when, and we don’t store it. Search and Ask work in English out of the box, and an optional free download adds 14 more languages to the phone itself. That download is language files, not the asking: Ask still needs the connection described above. It is on the App Store now, with Mac and more platforms next — get Fly Notes.

What is on-device voice transcription — does it work offline?