That seems too granular, as an additional 'uh' or 'and' would cause a sentence to be seen as different.
It might work well if you can first pair up the similar sentences from A and B, using word-level edit distance, or mapping to a lower-dimensional space using sentence embedding?
Can't help with the diff tool concept, but I built a search engine for talks, which semi-randomizes the results to force a variety in the search results.
This includes the older startup school talks, plus a few good software business conferences. For the next phase, I want to highlight talks that are most unique to the category (using word lists in the transcript/captions, not video diff)
Which makes me think: it would be interesting to have a diff tool for videos, but it diffs a representation of the content not specific video frames.
Customers might include all the students of Udacity, Coursera, EdX and Startupschool.