Just because it's possible doesn't mean it is clearly the answer. Is a transformer model truly likely to require less compute than current methods? We can't even run models like Veo 3 on consumer hardware at their current level of quality.
This was my thought as well. Why are we assuming they'd even have access to many preserved audio recordings knowing that they destroyed all computers? The Butlerisn Jihad was thousands of years before the events of the movie.