Transcribe a meeting on your own machine, not in the cloud

How a browser can transcribe speech at all

A speech model is a large file full of numbers that turns sound into words. For years these were too big and too slow to run anywhere but a data centre.

That is no longer true. The models got smaller, browsers got much faster at running heavy code, and an ordinary laptop can now transcribe an hour of talking without help from anyone.

Private Recap downloads that model once, keeps it on your machine, and feeds it the audio from your meeting.

Capturing the call, from a tab or a desktop app

First tell the recorder where the meeting is. A browser tab covers Google Meet and the web versions of Zoom and Teams; a desktop app covers the Zoom or Teams app on your computer. When you start recording, the browser asks what to share: the meeting tab, or your entire screen for a desktop app.

In that same window, make sure audio sharing is on. If a share arrives with no sound, the app says so and does not start, rather than recording a silent call and transcribing nothing. Only sound is recorded: when you share a screen, its picture is dropped the moment recording starts and never saved.

Leave the microphone toggle on if you want your own voice in the transcript as well as everyone else’s.

What happens on the first run

The first recap downloads about 300 MB of model files with the default models. That is a real download and the app says so before it starts, because a few hundred megabytes is not something to spend on somebody’s behalf.

Your audio, transcript and notes stay on your computer. The app goes online for three things only: the one-time model download, your account and billing if you choose to have one, and anonymous page counts.

After that download finishes you can transcribe on a plane.

How long it takes, and how good it is

Speed depends on your machine. A recent laptop handles a short call in a few minutes. An older one takes longer, and the tab has to stay open while it works.

Accuracy depends mostly on the audio. Clear speech through a headset transcribes well. Crosstalk, background noise and one room microphone at the far end of a table are harder, and that is true of every transcriber, wherever it runs.

Names and in-house acronyms are the usual source of mistakes. The transcript can be edited, and the notes are generated from it, so a minute of tidying before you generate pays for itself.

Limits worth knowing before you start

• Chrome and Edge can capture a meeting’s sound. Other browsers may be missing that, WebAssembly or local storage; the app lists what yours has.
• Desktop apps need Chrome or Edge on Windows, or on a Mac with macOS 14.2 or later and Chrome or Edge 142 or later.
• Free recordings stop at five minutes. Pro lifts that.
• Free uploads stop at five minutes too. On Pro an uploaded file can run up to an hour or 200 MB.
• The tab must stay open while a recording or a transcription is running.
• Nothing syncs. A meeting lives on the machine that recorded it, which is the point, but it does mean no colleague can search it.

From transcript to something you can send

A transcript on its own is rarely what you actually needed. The same pass that produces it also writes the notes, pulls out the action items and drafts the follow-up email.

Copy any part of it, export the lot, or regenerate one section that reads badly. What happens to it after that is between you and your own files.