For years, “transcribe it on your own computer” meant accepting garbage output. Two things ended that: OpenAI released Whisper as open weights, and Apple Silicon Macs shipped with enough GPU and unified memory to run its largest model in real time. Today an M-series Mac transcribes a meeting on-device, in the same accuracy class as cloud services, because it runs the same class of model.

The pipeline, end to end

A local meeting transcription setup does four things, all on your Mac:

Capture. macOS lets an app record system audio (what you hear, meaning the other side of a Zoom, Meet, or Teams call) alongside your microphone. No bot joins the call. The app taps the audio your Mac is already playing.
Speech-to-text. The app chunks the audio and feeds it through Whisper running on the Mac’s GPU. Whisper is a model with about 1.5 billion parameters, trained on hundreds of thousands of hours of multilingual audio. The large-v3 variant is what most cloud transcription is benchmarked against, and it supports 99 languages.
Cleanup. Raw speech-to-text output is accurate but ugly: fillers, false starts, broken punctuation, and the occasional mis-heard domain term. A small on-device LLM rewrites the raw transcript into clean prose. This pass is what makes local output read like a cloud service’s.
Write to disk. The result lands as a file you own. In Quietly’s case, that is plain Markdown in a folder you pick, readable by any editor and by Claude, Cursor, or Codex without an export step.

The app downloads the models once (a few gigabytes). After that, the whole pipeline runs without a network connection. Nothing uploads at any step. That is the entire privacy argument, reduced to architecture: there is no policy to trust, because there is no transmission to police.

Why Apple Silicon made this possible

Whisper large-v3 wants a GPU and several gigabytes of fast memory. On Apple Silicon, both come standard:

Unified memory lets the GPU address the same RAM as the CPU. There is no separate VRAM budget, so even a base M1 with 16GB can hold the model.
Metal gives the GPU general-purpose compute. Projects like whisper.cpp, a C++ port of Whisper tuned for Apple hardware, use it to hit real-time and faster transcription speeds.
Efficiency cores keep the machine usable while transcription runs in the background.

That is why this became a “Mac thing”: the hardware floor for running frontier-class speech models on-device arrived first, and most broadly, on Apple Silicon.

Is it as accurate as the cloud?

Much of cloud transcription is Whisper, or a model trained to beat it on the same benchmarks. When the model runs on your GPU instead of theirs, the words come out the same. Word-error rates for on-device large-v3 match its published, cloud-benchmarked numbers, because the same weights do the same math.

Local setups used to lose on presentation rather than recognition. Cloud services layer post-processing on top of the raw model output, and the on-device LLM cleanup pass restores that. Raw Whisper gives you “so um yeah i think we should uh ship the the api thing tuesday”. The cleanup pass gives you “I think we should ship the API change Tuesday.” Same information, readable note.

The DIY route

If you are comfortable in a terminal, you can prove all of this to yourself in ten minutes with an existing recording:

brew install whisper-cpp
whisper-cli -m ggml-large-v3.bin -f meeting.wav -otxt

That route is free, and for one-off files it works well. A DIY setup doesn’t give you the meeting-shaped parts: capturing both sides of a live call, knowing which meeting is which through your calendar, the cleanup pass, and notes that file themselves. Apps fill that gap. We compare the serious options, including the free, open-source ones, in the best local meeting note takers for Mac.

What you need

Apple Silicon (M1 or later; any tier works, and more GPU cores transcribe faster)
macOS 13+
A few GB of disk for the speech and cleanup models, downloaded once
No account, no API key, no network after that first download

If privacy is why you care about local transcription (NDA calls, hiring, legal, user research), the architecture above is what “private” should mean: the absence of a server, rather than a promise about one. More on that in private meeting transcription for Mac. If your starting point is “Zoom, but no bot in the call,” start with transcribing Zoom without a bot.