SubTitle Edit even tries to determine where best to "split" the caption.

I wonder how it does that ;) Might have explained my questions sooner.

So it uses an alligner which means it will check if a word corresponds with the audio. Much easier to do than having a machine learning algorithm like a neural network which tries to guess what's being said by having billions of training examples.

This way YouTube can allign transcripts to audio and figure out time codes. The probability of the next uttering being the word you are looking for is much higher than not knowing what's being said at all. Kuddo's to Google's neural networks!