in order to train doesn't it have to match audio output to a video of mouth movement?

Doesn't deep learning imply training on sample result?