Could you hack some sort of karaoke system? I mean, it ties word display to time, so you could then use that alongisde tying time to beats (DJ tools do this already) to link the words with the beats of the song. I would expect it to be manually generated but it'd give you a data set to work with for more popular stuff?

That said, I'm not sure how fine-grained karaoke systems get - whether they just display the lyrics for a whole bar of the song and linearly interpolate between the start and end, or whether they are a bit more intelligent than that.

Just a thought!