Transcribe a recording on your own computer with Whisper

Before anything technical, make sure everyone on the recording knows it's being recorded and transcribed. For a lecture, check your school's rules on recording classes. Running the model locally keeps the audio on your computer, but it doesn't replace consent.

Whisper's official repository says to install it with the command pip install -U openai-whisper, which uses Python's package installer. Its package listing names Python 3.8 through 3.13 as supported, newer than the README's older 3.8 to 3.11 line. It also needs ffmpeg, a free audio tool. On a Mac with Homebrew that's brew install ffmpeg, and on Ubuntu it's sudo apt update && sudo apt install ffmpeg. The README lists Windows options, too.

Open a terminal in the folder with your audio file and run whisper interview.mp3 --model turbo, swapping in your file name. The README says turbo is the default model and works well for English. The model file is saved to a cache folder on first use, so the first run takes longer.

Check: The folder now has transcript files, including a .txt for reading and .srt and .vtt caption files.

Whisper comes in six sizes. The README lists turbo as needing about 6 GB of video memory, the memory on a graphics card. If it's slow or crashes, try --model base or --model small. For English-only audio, the README says the .en versions tend to do better, especially tiny.en and base.en.

Speech-to-text guesses at words it can't hear clearly, and names, jargon, and numbers are the usual casualties. Listen back to every line you plan to quote or publish, and fix it before it goes into captions, show notes, or a study guide.