Hours of audio used to train OpenAI’s Whisper translation

Whisper was trained on 681,000 hours of audio from the web, with big variations between Indian languages