Hours of audio used to train OpenAI’s Whisper

Whisper was trained on 680,000 hours of audio from the web, with big variations between languages