Since it should be easy-ish to synthesize training data, I'd think combining this with deep learning could produce some good results.

Digitization, colourization, audio resampling, and other computer-based treatments of original recordings have taken over as restoration techniques from what are now archaic analogue-based techniques.