Textbook of AI

Abstract. OpenAI's Whisper paper. Trains a single encoder-decoder Transformer on 680,000 hours of multilingual, weakly-labelled web audio paired with transcripts of varying quality. The model handles speech recognition, translation and language identification across 99 languages in a single network and matches or exceeds specialised commercial systems on most benchmarks. Whisper is the canonical example of an encoder-decoder Transformer where the encoder and decoder serve clearly distinct roles, and the open-weights release made high-quality multilingual transcription a commodity capability.