As explained above, the process will take a lot longer as sounds get longer. Square the time. So if a 10 second sound is processed in 5 sec, a 20 second sound will take 25 sec, and a 40 sec sound will take 33 minutes… so the shorter the sound the much much more efficient the process will be.
The idea is not ‘for bufnmfcross’ but to remove duplicate sounds (so you have less sounds to process). The idea is to describe the sounds somehow (MFCCs might do) then map them in 2d with UMAP (the cleverest way we have to do dimension redux) then to use grid to force that map on to a 2d equidistant plane. Then, you skip boxes:
1 2 3
4 5 6
7 8 9
if you keep 1 3 5 7 9, or 2 4 6 8, you have a ‘downsampled’ dataset that is more representative (covers a wider space) than taking 1 2 3 4.
Does it make sense?