researchers-from-the-university-of-hamburg-propose-a-machine-learning-model-called-lipsound2-that-directly-predicts-speech-representations-from-raw-pixels