Excellent work. The results are pretty satisfying[0].
I wonder if it could be improved by relying on published lyrics instead of just video transcripts. I say that because the LMFAO song Party Rock Anthem[1] had some awkward moments when the karaoke display showed words that aren't actually part of the song, but are part of the video transcript. I suppose any song that has some sort of scripted intermission/narrative will have this type of problem.