Windows guide

Whisper on Windows without a terminal

Written 11 September 2026, based on ScreenBrief 0.7.8. Artflow Management.

Whisper is a speech recognition model that runs on your own computer. The classic route goes through Python, a package manager, ffmpeg on the PATH and graphics drivers. It does not have to. An app can ship all of that and fetch the model itself. Below: what is actually needed, which model to pick and how much room it takes.

Short answer

Install ScreenBrief from the Microsoft Store, pick a transcription model from the list and start the download. The media converter travels with the app, so ffmpeg is not a separate install. There is no Python step and no command line.

1. Install the app from the Store and walk through the short setup: language, transcription model, optionally an AI provider for summaries.
2. Pick a model. Setup finishes right away; you do not have to wait for the download before using the app.
3. Download the model now in Settings, or leave it to the first transcription, which fetches it on its own. The file size is shown before it starts and the download can be stopped.
4. Load a recording. Transcription runs on your computer and the result lands in the library.

What you do not have to install

• Python or a package manager. The speech recognition engine is compiled and shipped with the app.
• ffmpeg. The app carries its own copy and prepares the audio from a video file itself.
• CUDA drivers just in case. Transcription runs on the processor, and graphics card support is a separate, deliberate choice.
• The command line. There is no step where you retype a command from somebody's blog post.

That is the difference the question “Whisper without a terminal” is usually about. The model and the engine sit on your disk exactly as they would after a manual install; only the way they get there changes.

Which model to pick and how big it is

There are three Whisper models in GGML format. They differ in file size, processing time and quality on difficult material: noise, overlapping voices, rare names.

• Base, about 150 MB. The fastest one, basic quality. Good for a quick pass over a long recording and for weaker machines.
• Small, about 490 MB. The compromise between speed and quality, and the default choice in setup. Start here.
• Large v3 turbo, quantised, about 870 MB. The best quality and the longest processing time on a processor.

You can change the model later in Settings; downloaded files stay, so going back to a smaller model needs no second download. On a slow connection start small and swap later.

Where the model file comes from

The model is downloaded from the public whisper.cpp repository on Hugging Face, from one pinned repository revision rather than a moving branch. Once it arrives, the app checks the file's SHA-256 checksum against the value stored in the program.

So bytes swapped under a familiar file name do not pass unnoticed, and an interrupted download is not mistaken for a complete model. You can stop the download with a button; the unfinished file is cleaned up.

When the graphics card helps

On a fresh Windows install the transcription device is set to “Automatic (recommended)” and Settings shows a “GPU engine” section. A “Check now” button runs the same probe that processing uses and says whether the card can be used and why an install is or is not available.

There are two engines. Vulkan covers NVIDIA, AMD and Intel cards. CUDA is for NVIDIA cards and needs about 2.8 GB of verified files installed from the app package; after processing restarts, the “Active engine” field shows what actually runs.

It works without a graphics card too, just slower. A large model on a processor can take a long time, and the app says so while it works.

What stays on disk

The details are in the guide on transcription without the cloud. If the summary should be produced on your computer as well, see the Ollama guide.

When this is the wrong choice

• When you need Whisper inside a script or a server pipeline. That is exactly what a command-line install is for, and nothing replaces it.
• When you want to hand-tune decoding parameters on every run. The app picks them for you.
• When you work outside Windows. This guide describes the Store version for Windows 10 and 11.

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