Can AI transcribe meetings without the internet?

Yes. Modern speech-recognition models such as Whisper run entirely on a Mac's own CPU and GPU. Once the model is downloaded, transcription and summarization happen on-device with no network connection, so no audio is uploaded to any server.

Is on-device transcription as accurate as cloud transcription?

For most meetings, yes. On-device tools use the same families of open-source models (such as OpenAI's Whisper) that many cloud services are built on. Accuracy depends on audio quality, accents and the model size chosen, not on whether the processing happens locally or in the cloud.

Does on-device AI drain my battery or slow my Mac?

Capturing audio is lightweight. The heavier transcription and summarization work runs after the meeting at low priority on Apple Silicon, so it uses spare capacity and does not interrupt your call.

How on-device meeting transcription actually works

"On-device" and "local AI" get used as marketing words, so it's worth seeing what they actually mean. Turning a two-hour call into a tidy set of notes takes four steps — and on a modern Mac, every one of them can happen on the machine in front of you, with nothing uploaded to anyone's cloud.

Step 1 — Capturing the audio

A meeting has two audio streams: your microphone, and whatever is coming out of your speakers (the other people on the call). A cloud notetaker usually solves this by sending a bot into the meeting to record the mixed room. A local tool instead captures both streams directly on your Mac — your mic through the normal microphone permission, and the system output through a small audio component — and mixes them into one recording. No bot ever appears in the participant list, because nothing joins the call.

Step 2 — Speech recognition (speech → text)

The recording is handed to an automatic speech recognition (ASR) model. The most widely used open one is OpenAI's Whisper, which many cloud services are also built on. The key point: Whisper is a file you can download and run yourself. Projects like whisper.cpp compile it to run efficiently on Apple Silicon's CPU and GPU, so a Mac can transcribe audio with no network connection at all.

This is the part people find surprising: the intelligence isn't in a data centre, it's in a model file sitting on your disk. Once it's downloaded, you could pull the Ethernet cable and transcription still works.

Step 3 — Figuring out who said what (diarization)

A raw transcript is a wall of text. Speaker diarization splits it into turns — "Speaker 1 said this, then Speaker 2 said that" — by clustering voices that sound like the same person. This also runs locally. It's never perfect (two similar voices, or lots of crosstalk, will trip any system), but it's what turns a transcript into something that reads like minutes.

Step 4 — Summarizing (text → notes)

Finally a small language model reads the transcript and writes the summary, the decisions, and the action items. Cloud notetakers send your transcript to a large hosted model for this. On-device tools use a compact local model — QuietScribe uses Google's open Gemma — that is small enough to run on a laptop while still being good at "read this and pull out the to-dos." The trade-off is honest: a giant frontier model in the cloud may write a slightly slicker paragraph, but a local model keeps the entire transcript on your machine, which for confidential calls is the whole point.

So what's the catch?

A one-time download. The models are a few gigabytes; you download them once during setup, then everything is offline.
It needs decent hardware. On-device AI is why these tools target Apple Silicon — the Neural Engine and unified memory make local models practical on a laptop.
Processing takes a moment. The summary appears after the meeting rather than streaming from a server farm. In exchange, nothing about the meeting ever leaves your Mac.

Why the location of the processing matters

When transcription happens in the cloud, your meeting audio and transcript exist as a copy on a vendor's servers — subject to their retention policy, their security, their data-training settings, and any subpoena or breach that reaches them. When it happens on-device, there is simply no second copy to leak. For lawyers, doctors, therapists, HR, founders discussing unreleased work, and anyone under an NDA, that difference is the product.

QuietScribe runs all four of these steps on your Mac: it captures both sides of the call, transcribes with whisper.cpp, diarizes locally, and summarizes with Gemma — producing Markdown notes in Obsidian, Apple Notes or Notion. Flat $10/month for one Mac, unlimited minutes, no cloud copy.

Try QuietScribe — free private beta