The scripts that pack audio files into HDF5 files from the PANN paper are quite useful. However, one upgrade I sought out was running the packing in parallel since the AudioSet dataset is most usable when separated into a number of HDF5 files, not just one giant file.
I realize now that multiprocessing for this application may have been excessive... However, it has significantly reduced the packing time.
This is a modified version of the original repo which allows for one to pack the HDF5 files in parallel. Here is a screenshot of it in action: