Sound samples can be found in "low_bitrate_example.mp3" and "reconstructed_example.mp3". Not too shabby!
Copy a bunch of good-sounding mp3's into data/mp3_high
Run the convert.py tool to re-encode them into garbage-sounding 64kbps MP3s
Run restorer/train.py to train a GAN that tries to reconstruct portions of the low-bitrate MP3s in the frequency domain.
Here's an example reconstruction of the spectrogram:
Writeup coming soon...