Whisper on Windows, without Python or the command line
OpenAI's Whisper is one of the best speech recognition models you can run yourself. Running it the usual way means Python, ffmpeg and a terminal. Transcribe packages it as a Windows app, with the models included and an editor around them.
What Whisper is
Whisper is a family of speech recognition models that OpenAI released in September 2022 under the MIT licence, code and weights included. It was trained on 680,000 hours of multilingual audio from the web, which is why it copes well with accents, background noise and many languages. It writes punctuation and capital letters, detects the language, and can translate speech into English.
The models come in sizes from tiny to large. Larger models make fewer mistakes and need more time and memory. Large-v3 arrived in late 2023; large-v3-turbo, in 2024, cut its decoder from 32 layers to 4, which makes it several times faster while keeping most of its accuracy.
The usual way, and its hurdles
The official route runs through Python:
pip install -U openai-whisper
# also needed: ffmpeg, installed and on the PATH
whisper interview.mp3 --model small --language EnglishIt works, and it is free. The hurdles are the setup (a Python environment, ffmpeg, and for speed an NVIDIA graphics card with CUDA), models that download on first use, and the result: text and subtitle files with no editor and no way to hear a line again. whisper.cpp, a fast C++ port, is lighter, but it is still a command-line program with models to fetch.
What Transcribe adds around Whisper
| Command-line Whisper | Transcribe - Offline | |
|---|---|---|
| Install | Python, packages, ffmpeg | One app from the Microsoft Store |
| Models | Downloaded on first use | Included: Whisper small and large-v3-turbo |
| Graphics acceleration | CUDA, so NVIDIA cards | Vulkan, so AMD, Intel and NVIDIA graphics |
| Without a graphics card | Slow | Works: Standard at about 3× real time on our test laptop |
| Checking the result | A text editor and a media player | An editor tied to the audio: click a line to hear it |
| Long recordings | One run; start again if it stops | Background jobs that continue after a restart |
| Speakers | Not included | Speaker detection (Pro), naming by hand (Free) |
| Output | txt, srt, vtt, tsv, json | Copy, txt, srt, vtt; Word and PDF with Pro |
| Cost | Free | Free with Whisper small; Pro for more |
Which Whisper models Transcribe uses, and why
Standard is Whisper small, quantized to keep it compact. It is the smallest model whose accuracy we were happy to call standard: on our test set it got 2.0% of words wrong on clean audio, and it ran at about 17× real time on a laptop's integrated graphics. Tiny and base are faster, but make too many mistakes to save you time.
Accurate is Whisper large-v3-turbo, also quantized. It got 1.1% of words wrong on the same audio and held up better in noise, at about 2.5× real time on the same laptop, an hour of audio in about 24 min. Full large-v3 would be slower still for little gain.
Both run in whisper.cpp, which uses the graphics chip through Vulkan when the driver supports it, and the processor otherwise. The full measurements.
What Transcribe does not do
- No translation. Whisper can translate speech into English; Transcribe transcribes in the language that was spoken.
- No custom models. The two included models are the ones we test and measure.
- Windows only. The command-line tools run on macOS and Linux too.
Questions
Is Whisper free to use?
Yes. OpenAI released Whisper's code and model weights under the MIT licence. Transcribe's Standard model, Whisper small, is free without limits in the app.
Do I need an NVIDIA graphics card for Whisper on Windows?
Not with Transcribe. It uses Vulkan, which AMD, Intel and NVIDIA graphics support, and falls back to the processor. On our test laptop's integrated AMD graphics, Whisper small ran at about 17× real time.
Which Whisper model is best?
For everyday use, small is a good balance of speed and accuracy. For noise, accents and names, large-v3-turbo is the better choice and much faster than the full large-v3.
Does Whisper work offline?
Yes. Once the model is on your PC, Whisper needs no internet. In Transcribe the models are installed with the app, so it works offline from the first run.
Your recordings stay on your PC.
Free to download. Transcribes without an account, without uploads and without the internet.
Coming soon to the Microsoft Store · Windows 10 and 11