OFFLINE TRANSCRIPTION

Offline transcription on a Mac

"Offline" gets used loosely. Here is what it means mechanically, what a Mac can actually do without a connection, and how to tell a genuinely local app from one that merely starts locally.

What the Mac is doing

Apple Silicon Macs carry a Neural Engine, a block of silicon built for running machine-learning models. Speech recognition is one of the things it is good at. A transcription model is a file on disk; running it is arithmetic on that file plus your audio. Nothing about the arithmetic requires a server. The reason most transcription happens in the cloud is that it was easier to sell and easier to update, not that a laptop cannot do it.

Speaker separation, usually called diarization, is a second model doing a related job: deciding where the voice changes and cutting the transcript there. It runs on the same hardware.

The parts that do touch the network

The models have to arrive somehow. Every on-device transcription app downloads them the first time, usually a few hundred megabytes. After that the files are local and transcription works with the wifi switched off. A signed app is also likely to check whether a newer version exists, which is a request carrying a version number and nothing else.

So "offline transcription" describes where the transcribing happens. It is not the same as an app that never opens a socket, and an honest one tells you which requests it makes and lets you switch off the ones that are optional. Be wary of a page claiming "100% offline" with no asterisk at all: the models did not appear by magic, and the claim is usually a sign nobody checked.

For the record, Memory Notch makes exactly two: the models, downloaded once and fetchable ahead of time from Settings, and a daily version-number check in signed releases that Settings can switch off. No audio, transcript or note is ever uploaded.

How to check an app is really local

• Turn the wifi off and record something. The bluntest test there is.
• Read the privacy policy for retention language. A local app has nothing to retain.
• Ask whether there is an account. Sign-in usually implies a server holding something.
• Look for a data-residency page. If the vendor has to tell you which region your data lives in, it lives somewhere.

What you give up

Cloud transcription runs larger models on better hardware, and it shows on hard audio: strong accents, people talking over each other, a bad room. On-device models do well on clear conversation and are improving quickly. A custom vocabulary helps more than people expect, because most of what reads as an error is a proper noun the model had never seen. Memory Notch takes a list of names, companies and jargon from vocabulary.md and matches them against the audio instead of guessing phonetically.

You also give up the team features. Shared workspaces, CRM sync and cross-meeting analytics are cloud products by nature. That is the trade, stated plainly.

Which Mac apps run locally

Fewer than the marketing suggests. Hyprnote and Meetily are open source and genuinely on-device. MacWhisper transcribes files on-device very well. Granola captures locally but transcribes in the cloud. Otter, Fireflies and Fathom are cloud products. The full comparison sorts all of them by where the audio goes.

Memory Notch records from the notch, transcribes and labels speakers on the Neural Engine, and writes each recording into ~/memory as plain Markdown. Try the demo.