How do you install Whisper on Windows?

Whisper is OpenAI's open speech recognition model, and its official README installs it as a Python package with pip ¹. On Windows that means four steps in a terminal: install Python, install ffmpeg, run the pip command, then test it on a file. If you do not want to use a terminal, skip to the Buzz section below.

Install Whisper in four steps

1. Install Python 3.8 to 3.11. The README says the codebase is "expected to be compatible" with that range ². Newer Python releases may work, but this range is the one the project names. Make sure Python is on your PATH so the terminal can find it.
2. Install ffmpeg. Whisper requires the command-line tool ffmpeg on your system ³. If you already have Chocolatey, run choco install ffmpeg in a terminal ⁴. Then run ffmpeg -version to confirm the tool works.
3. Install Whisper. In the same terminal, run pip install -U openai-whisper. The -U flag upgrades an existing install to the latest release.
4. Transcribe a test file. Put an audio file in the folder you are working in, then run whisper audio.mp3 --model turbo. The README's example transcribes speech in audio files with the turbo model ⁵. This is the same form, with one file.

If the install fails

One failure the README warns about is a missing wheel for tiktoken, a package Whisper depends on. The README says you may need Rust installed in that case ⁶. Install Rust, then run the pip command again. If the test command cannot find ffmpeg, open a new terminal after installing it, so the new path takes effect, then check again with ffmpeg -version.

Choose a model your memory can hold

The model you pick sets how much memory Whisper needs, and larger models need more. Check the installed memory in Windows before you choose. If your PC has less memory, use a smaller model, such as --model small, in the test command. Our guide to offline speech to text on Windows lists the memory figures for the tiny, small and large models.

Want a window instead of a terminal? Use Buzz

Buzz is a separate app, and it has a Windows installer. The installer is not code-signed, so Windows will show a warning when you install it. The README explains how to continue past that warning ⁷. Download the installer only from the project's own download page.

Buzz also has a pip package, and its README asks for a Python 3.12 environment for that route ⁸. That differs from the Python 3.8 to 3.11 range for Whisper itself, so keep the two installs in separate environments if you use both.

Which route should you pick?

• You are comfortable in a terminal and want the official package: use the pip route above.
• You want a window and no terminal: use the Buzz installer.
• You want live dictation into any app, not audio files: Whisper's command line is the wrong tool. Read whether there is a Whisper app for Windows and which open-source dictation apps run on Windows instead.

Test it on your own voice

A short file tells you whether the install works. It does not tell you how accurate the model is on your speech. Record two minutes of your normal writing, transcribe it, and read the result against what you said. Count the wrong words, the missing words and any invented words. The word error rate is the usual measure for these counts, and our guide to running Whisper locally explains how to get the most from a local install.

If Whisper works but Windows dictation into your apps does not, that is a separate question. Our answer on whether Windows voice typing works offline explains why Win+H behaves differently. For the whole subject, see our guide to offline speech to text on Windows and the Windows dictation overview.