On-device transcription: how it works and what it's for

In short

On-device transcription means the speech recogniser runs on the computer that holds the recording. The audio is turned into text by your own processor, and nothing is sent to a server to be transcribed.

On a Mac running macOS 26, Apple provides such a recogniser as part of the operating system, and apps can use it without shipping a model of their own.

This guide explains what happens when a Mac transcribes speech on the device, what that changes for privacy, where it falls short, and how a meeting notetaker can build on it.

Two places a transcript can be made

Every transcription tool has to answer one question: where does the audio go to become text?

	Cloud transcription	On-device transcription
Where the audio goes	To a speech service run by the vendor or a provider it uses	Nowhere; it stays on the computer
Needs a connection	Yes, for every recording	Only to download the model once
Who else holds the audio	The service, for as long as its terms say	Nobody
Limits	The service's languages, pricing and retention	The device's languages, hardware and model size

Cloud recognisers can be very large and can be updated without touching your computer. On-device recognisers have to fit on a laptop and run fast enough to keep up with speech, which used to mean a clear drop in accuracy. That gap has narrowed, but the trade is still there: size, languages and hardware on one side, where your audio goes on the other.

What "on-device" means on a Mac

macOS has had an on-device option for years. Apple's older Speech API, SFSpeechRecognizer, has a setting called requiresOnDeviceRecognition that stops a request from sending audio over the network. Apple's own documentation adds that "on-device requests won't be as accurate", and that for some languages the older recogniser needs an internet connection at all.

In macOS 26, Apple added a new API, SpeechAnalyzer, with a new transcription module, SpeechTranscriber. Apple describes the model behind it as "good for long-form and distant audio, such as lectures, meetings, and conversations", and says it runs on the device. It is the same model Apple uses in its own apps, including Notes. Apple's SpeechAnalyzer covers the API in more detail.

A few details matter to anyone relying on it:

The model belongs to the system, not the app. macOS downloads it from Apple's servers, keeps it, updates it and shares it between apps. An app asks for a language; the system fetches what is needed once.
Nothing is transcribed over the network. The download is the model coming to your Mac. The audio does not go the other way.
It works while you talk. The transcriber gives quick guesses as words are spoken ("volatile" results) and replaces them with settled text a moment later ("finalized" results). That is why live transcripts often show a phrase changing as the sentence finishes.
Hardware and languages are limited. Apple says SpeechTranscriber is available "with certain hardware requirements" and gives apps a check for whether the current device supports it.

What it changes for privacy

When transcription runs in the cloud, a meeting's audio is copied to at least one other organisation. That organisation becomes part of every question about the meeting: how long it keeps the file, whether it trains on it, which country it sits in, and whether it is covered by your contract with the vendor. Under GDPR it is a processor you have to account for.

On-device transcription takes that step out of the chain. The audio exists in one place, your disk, and the transcript is written next to it.

It is worth being exact about what it does not change:

The recording still exists. It is on your laptop, with whatever protection your laptop has. Disk encryption and your own backups decide who else could read it.
The text may still travel. If you ask an AI to summarise the meeting, the transcript text goes to whoever runs that AI. On-device transcription keeps the audio at home, not necessarily the words, and it does nothing to make a summary correct: can you trust an AI meeting summary? covers checking one.
The law still applies. Whether you may record, and what you have to tell people, does not depend on where the transcript is made. Recording meetings under GDPR goes through what data-protection law asks.

What it is good for

On-device transcription suits situations where the audio is the sensitive part and a network is optional:

Meetings you would rather not upload: client calls, interviews, conversations under an NDA.
Work without a connection: on a train, on a plane, or on a network that blocks unknown services. Transcribing a meeting offline shows what works with no network at all.
Organisations that restrict vendors: every service that receives audio needs approval, and a recogniser that is part of the operating system needs none of its own.
Long recordings: there is no upload to wait for and no per-minute bill for transcription.
Access: a transcript on the screen during a call and afterwards helps people who are deaf or hard of hearing, or who find it hard to listen and write at once. Live captions for meetings on a Mac, meeting notes for people who are hard of hearing, note-taking with dyslexia and transcripts as a reasonable accommodation cover this.
Finding things later: with the transcript tied to the recording, finding a moment in a meeting recording is a search rather than a scrub through the audio.

Where it falls short

Be clear about the limits before relying on it.

Languages. Apple's new transcriber covers far fewer languages than its older dictation engine. As of September 2026 it reported ten on macOS 26.6. If you meet in a language outside that list, it will not transcribe it. See which languages a Mac can transcribe on the device, and for particular languages, transcribing a meeting in French, Spanish, German, Japanese or Chinese. A meeting that switches between languages is harder again; multilingual meeting transcription explains why.
Hardware and macOS version. SpeechAnalyzer is new in macOS 26, so older versions of macOS do not have it at all.
The hard words. Names, numbers, product names and jargon are where every recogniser goes wrong, local or not. Why meeting transcripts get names and numbers wrong explains why, word error rate explains how accuracy is measured and why vendor figures rarely compare, and getting better meeting transcripts covers what you can change in the room.
Live is not final. A recogniser working in real time commits to words before it has heard the end of the sentence. A pass over the whole recording afterwards can do better, which is the subject of live versus post-meeting transcription.
Who said what. A recogniser turns speech into words; it does not know who is speaking. That is a separate step, covered in speaker diarization explained; putting names to the voices afterwards is covered in how to label speakers in a meeting transcript.

How to check whether an app really transcribes on the device

A description saying "on-device" or "private" is not evidence. These are:

Turn the network off and try it. Switch Wi-Fi off, record a minute of speech and see whether a transcript appears. If the app needs a connection to transcribe, the audio is going somewhere.
Read the privacy policy for the word "audio". It should say plainly whether audio is uploaded, and if so to whom.
Look for a transcription provider in the processor list. A vendor that sends audio to a speech service has to name it somewhere.
Separate transcription from AI notes. An app can transcribe on the device and still send text to a language model for summaries. Both facts should be stated.

How Notey uses it

Notey is a meeting notetaker for Macs with Apple silicon running macOS 26 or later. It records your microphone and the sound your Mac is playing as two separate tracks, and transcribes both on the Mac.

Live transcription uses Apple's on-device speech recognition, in the ten languages it supports. Words appear as they are spoken; a phrase still being revised is shown lighter until it settles.
Your side and theirs come from the tracks. Each line is attributed to the track it arrived on, so the transcript knows "You" from "Them" without learning anyone's voice.
After an English meeting, Notey reads the recording again on your Mac with two stronger on-device recognisers and keeps the words they and the live transcript agree on. They are about 1.2 GB in total, fetched once in the background, never run or downloaded during a recording; a meeting you have changed is left as recorded, and the whole thing can be switched off. Why Notey re-transcribes a meeting after it ends explains it.
No account and no network are needed to record, transcribe, search and read meetings back. Audio is never uploaded to be transcribed.
AI notes are separate. If you ask for a summary or other notes, the transcript text, not the audio, is sent to Notey's service and on to OpenAI as a processor. The privacy policy sets out exactly what is sent and when.

If you want to try it, download Notey and record a call with the Wi-Fi off.

Frequently asked questions

Is on-device transcription as accurate as cloud transcription?

It depends on the engine and the audio, and there is no single answer. Apple's newer on-device model was built for long-form audio such as meetings, while Apple's own documentation for its older recogniser says on-device requests "won't be as accurate" as server ones. Test on your own recordings.

Does on-device transcription need the internet?

Not to transcribe. The speech model has to be on the Mac first, and macOS downloads it from Apple once per language if it is not already installed. After that, transcription works with no connection.

Does on-device transcription work on an Intel Mac?

Apple says its new transcriber has "certain hardware requirements" and gives apps a way to check. Notey requires Apple silicon and macOS 26 or later, so it does not run on Intel Macs.

Which languages can a Mac transcribe on the device?

Apple's SpeechTranscriber reported ten languages on macOS 26.6 in September 2026. Which languages a Mac can transcribe lists them with their regional variants.

No. It removes the transcription service from the list of places the audio goes, which is a real simplification, but you still need a lawful basis, a notice and a retention period.