How do you use Whisper to transcribe audio?

Install the openai-whisper package and the ffmpeg tool, then run one command on your audio file. Whisper is OpenAI's open speech recognition model. Its code and model weights are released under the MIT License ¹, so you can run it on your own computer for free. The setup takes a few minutes in a terminal. The first real test should be a short file, and the transcript should be checked against the recording before you rely on it.

The steps below follow the README in OpenAI's repository. For the full decision between routes, read the guide to transcribing with Whisper, ChatGPT or Gemini, which sits in our transcription overview.

Whisper takes four steps: install, add ffmpeg, transcribe, then check

1. Install the package. In a terminal, run pip install -U openai-whisper ². The -U flag installs the latest release, or updates an existing one.
2. Install ffmpeg. The package also needs the command-line tool ffmpeg on your system ³. The README says it is available from most package managers, so use the one your system already has.
3. Transcribe your file. Run whisper your-recording.mp3 --model turbo. The README's example uses the turbo model to transcribe speech in audio files ⁴.
4. Check the output. Open the transcript and listen to a few passages against the recording. Look for wrong names, numbers and terms first.

If the first run fails, check step 2 first, because the package depends on ffmpeg being on your system.

Choose the model that your computer's memory can hold

The turbo model needs about 6 GB of video memory, according to the README's model table ⁵. The large model needs about 10 GB ⁶, so on a computer with less memory it is the first to fail. These are approximate figures, and they are for a graphics card. A computer without a suitable graphics card runs the model on its processor, which is slower and uses system memory instead.

If the run is slow, or it stops because the computer runs out of memory, switch to a smaller model with the --model option. Test your own recording with each size you consider, because the table does not measure accuracy for your audio. For a longer comparison, read which Whisper model is best.

Whisper can invent text in silence, so read the output beside the audio

Whisper can produce text that the audio does not contain. In a 2024 study, roughly 1 percent of Whisper's transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ⁷. The study tested a 2023 version, so newer releases may behave differently.

The risk is highest in quiet gaps, music and background noise. Before you transcribe a long file:

• Trim long silences and music-only passages from the start and end.
• Read the output at each long pause. A phrase that repeats, or one unrelated to the topic, is a warning sign.
• Re-run only the doubtful section on its own, then compare the two versions with the recording.

Some tools also run a voice activity detection step that marks where speech is. Check which parts that step removed, because a wrong cut can hide speech as well as silence.

Accuracy depends on the language, so test the language you use

Accuracy varies from one language to another. A word error rate figure for English does not tell you how well Whisper will do in another language, or on your speakers and recording quality. Run about ten minutes of your own audio through the model you plan to use, then read the transcript against the recording.

If you want fewer setup steps, try faster-whisper or a Whisper app

Unlike openai-whisper, the faster-whisper project does not need FFmpeg installed on the system ⁸. It is a separate project with its own command-line tool, so you do not need the openai-whisper install steps above.

A Whisper app with a window is another option. Check whether it runs the model on your computer or sends the audio to a server, since a server-based app sends your recording off the device. For a one-off file, the audio to text tool transcribes in your browser with no install. For a Mac, offline dictation on Mac covers local speech recognition in more detail.

Related questions

• Can ChatGPT transcribe audio files?
• Can Gemini transcribe audio?
• Can NotebookLM transcribe audio?
• Install Whisper on Windows
• Can Whisper run without a GPU?

For a terminology check, see automatic speech recognition, the technology Whisper is built to do.