They've trained and are open-sourcing a neural net called Whisper that approaches human level robustness and accuracy on English speech recognition.