Can you transcribe audio locally on a Mac?

Yes. Apple Silicon Macs can run speech recognition models directly on the machine, with no audio sent anywhere. The model file is downloaded once, then transcription works with the network off. You need a few gigabytes of free disk space and an M-series Mac for reasonable speed.

Is local transcription as accurate as cloud transcription?

For clear English speech on a decent microphone, on-device models are close enough that most people do not notice a difference. Cloud services still have an edge on heavy accents, poor audio, and languages beyond English, because they can run much larger models on server hardware.

How much disk space do local transcription models need?

Expect a few gigabytes. Private Notetaker downloads roughly 640MB for the transcription model and around 4GB for the summarization model. You download them once and they stay on the machine.

Does local transcription work offline?

Once the models are downloaded, yes. That is the practical test of whether a tool is really local: turn off wifi and try to transcribe. If it still works, the processing is genuinely happening on your machine.

How do I know a transcription app is not uploading my audio?

Turn off the network and use it. Beyond that, check whether the app asks for a login before it will transcribe, and read what the privacy policy says about retention rather than what the homepage says about privacy. A policy that promises deletion still means the audio was uploaded.

← All articles

Transcription
Local AI
Mac

Local AI transcription on a Mac: what actually runs on your machine

How on-device transcription works on a Mac, what the models cost you in disk and battery, and how to check whether a tool that claims to be local really is.

Ben Claybrook 28 July 2026 5 min read

If you have looked for a way to transcribe meetings on a Mac without shipping the audio to someone else’s servers, you have probably noticed that almost every tool says it is private and almost none of them explain what that means. Some encrypt the upload. Some promise to delete the recording afterwards. A much smaller number never upload anything at all.

This is a practical guide to the third category: what on-device transcription actually involves on a Mac, what it costs you in disk space and battery, and how to test a tool’s claims rather than trusting them.

What “local” actually means

Transcription has two expensive steps. First, speech has to be converted to text. Second, if you want a summary or action items, a language model has to read that text and write something useful.

A cloud tool sends your audio to a server for step one and your transcript to a server for step two. A local tool runs both on your Mac using models that live on your disk.

The distinction matters because the two designs fail differently. A cloud service with a strong privacy policy is still one breach, one subpoena, or one policy change away from exposing your recordings. A tool that never uploads has nothing to expose. That is not a stronger promise. It is a different architecture.

Why this became possible on a Mac specifically

Apple Silicon changed the maths. The unified memory on M-series chips lets a model access a large pool of memory without copying data back and forth to a separate graphics card, and the neural engine handles the kind of work speech recognition does. Models that needed a server three years ago now run on a laptop at usable speed.

The practical result is that a modern Mac can transcribe an hour of audio in a few minutes without touching the network, and can summarize the result immediately afterwards.

What it costs you

Local processing is not free. It trades a subscription for resources on your machine.

Disk space. Private Notetaker downloads about 640MB for transcription and about 4GB for summarization. That is a one time download, and we state the real numbers during setup rather than hiding them, because 4GB is a decision and you should get to make it.
Battery and heat. Transcribing a long recording will spin the fans and draw power. It is a burst, not a constant drain, but it is noticeable on battery.
Time to first use. You wait for a multi-gigabyte download before the first transcription. Cloud tools feel faster to start because the model is already sitting on someone else’s hardware.

If those trade-offs are not worth it for your work, a cloud tool is a reasonable choice. The point is to make the trade knowingly.

How do you verify a tool is really local?

Four checks, in order of how much they tell you.

Turn off the network and use it. This is the only test that cannot be argued with. If transcription still completes with wifi off, the processing is on your machine. If it hangs or errors, it was never local.

Check whether it needs an account to work. Tools that process on servers need to know who you are before they will do anything. A tool that transcribes before you have logged in is doing the work locally.

Read the retention language, not the privacy language. “We do not train on your data” and “we delete recordings after 30 days” both confirm that your audio was uploaded. The homepage word to look for is not private. It is on-device or local.

Look at whether a bot joins your calls. Meeting assistants that appear as a participant in Zoom or Meet are receiving the stream on a server. That is a visible, unambiguous signal about where the audio goes.

Where Private Notetaker sits, including what it does not do

Private Notetaker records, transcribes, and summarizes on your Mac. It captures system audio natively through macOS, so it records both sides of a call without a virtual audio driver and without joining the meeting as a participant. The transcript, the summary, the decisions and the action items are all produced on the machine. The app works with the network off, and there is no account required to take notes.

Being straight about the limits, because they are real:

Windows is a beta and it is unsigned. Windows SmartScreen will warn you and you have to click through it. The Mac build is signed and notarized; the Windows one is not yet.
English only at the moment.
No cloud sync and no team features. Your notes exist on the machine that recorded them. That is the design, but if you want a shared workspace where colleagues can read your meeting notes, this is the wrong tool.
Cloud services still win on hard audio. Heavy accents, several people talking over each other, a bad room. Server-side models are larger and they handle that better.

If your meetings are covered by a confidentiality obligation, or you work with clients whose information you would rather not hand to a third party, the architecture is the argument. If you mostly need a shared team notepad, it is not.

You can see pricing and try it, or read how it compares to the cloud tools people usually switch from.

Frequently asked questions

Can you transcribe audio locally on a Mac?: Yes. Apple Silicon Macs can run speech recognition models directly on the machine, with no audio sent anywhere. The model file is downloaded once, then transcription works with the network off. You need a few gigabytes of free disk space and an M-series Mac for reasonable speed.
Is local transcription as accurate as cloud transcription?: For clear English speech on a decent microphone, on-device models are close enough that most people do not notice a difference. Cloud services still have an edge on heavy accents, poor audio, and languages beyond English, because they can run much larger models on server hardware.
How much disk space do local transcription models need?: Expect a few gigabytes. Private Notetaker downloads roughly 640MB for the transcription model and around 4GB for the summarization model. You download them once and they stay on the machine.
Does local transcription work offline?: Once the models are downloaded, yes. That is the practical test of whether a tool is really local: turn off wifi and try to transcribe. If it still works, the processing is genuinely happening on your machine.
How do I know a transcription app is not uploading my audio?: Turn off the network and use it. Beyond that, check whether the app asks for a login before it will transcribe, and read what the privacy policy says about retention rather than what the homepage says about privacy. A policy that promises deletion still means the audio was uploaded.