Guide

Meeting transcription where the recording never leaves your computer

Written 3 September 2026, based on ScreenBrief 0.7.7. Artflow Management.

A recording can be transcribed entirely on your own computer. You hand the file to an app sitting on your own disk, so the recording goes nowhere: no account, no upload to a server. Here is what that looks like in practice, what does still leave the machine, and when you should pick something else.

The short answer

You need software that ships a speech recognition model and runs it locally rather than on someone else's server. ScreenBrief works that way: you point it at an audio or video file on your own disk, or record your screen, and transcription runs on your own CPU or graphics card. No sign-in, no account, and the file is never sent to a server.

1. Install the app and open it. It does not create an account.
2. Point the app at a recording already on your disk, or record your screen with system audio.
3. The first transcription downloads the speech model. That is the only moment anything is fetched from the internet.
4. The transcript appears with speaker labels and timestamps, and the recording stays in a folder on your own disk.

What stays on disk, and what still leaves

This is the part most descriptions get wrong by omission. The honest split looks like this:

The last two items on the right are settings you pick, not defaults. Out of the box transcription is local, and summaries wait until you name a provider.

Which model gets downloaded, and how big it is

The model is the one thing that has to be fetched, so it is worth knowing what you are agreeing to. There are three, and the most accurate one is on by default:

The download happens once and the model stays on disk. Dropping to a smaller model costs some accuracy, mostly on names and acronyms.

How long it takes on your hardware

There is no honest answer in minutes, because it depends on your CPU, your graphics card and the length of the recording. Instead of trusting someone else's benchmark, measure it once: run the default model on one short recording and see how long it took. That single measurement is what tells you whether to drop to a smaller model or hand transcription to an external provider.

Your graphics card is used if you have one. Which chips are covered is in the FAQ on the home page.

When local transcription is the wrong choice

Honestly, it happens. Three situations where something else is the better answer:

• You have many long recordings, need results immediately, and the machine is slow. API transcription will be faster, at the cost of the audio leaving.
• You work on a machine where you cannot store a few hundred megabytes of model weights.
• You need a contractual guarantee from a service provider, not just the fact that the file never leaves the computer.

The transcript is only the start

A transcript is rarely the goal. The same recording also produces summaries, the decisions that were made, the tasks that came out of it, and a chat that cites the second of the recording behind every answer. What exactly comes out of one recording is on the home page, and the terms and price are in the pricing section.

This describes ScreenBrief 0.7.8. Details of how data is handled: privacy policy.