Ask HN: Could deep learning be used to restore old musical recordings?

Let's say we limit the problem to old recordings of classical pianists from the early 20th century, could a deep learning algorithm be trained on pristine recordings of piano tones (or entire pieces) to restore the quality of an old recording? I mean more than just removing pops, scratches and surface noise -- I mean, adding a dynamic range to the recording that was never there, but would have been there if it had been recorded with modern technology.