ffmpeg -y -i "input.wav" -af aresample=32000:cutoff=1 "output.wav"

 

it is not such type of filters as you might think, but of course a proper algorithm does some filtering.

modern DAWs are all very good, yet not identical.


an ideal SRC algorithm goes over the least common multiple.

for example the LCR of 96000 and 44100 is 14112000.

so if you want to export a 44100 file from a 96000 file, you first make 147 copies of every sample, and then press 320 of these samples into 1 sample of the target file.
the filtering is very easy and basically only consists of averaging these 320 samples, and the result is close to perfect. it is just very slow to process...

 

I'd like to add that what you said is true in principle for non-integer resampling, but that upsampling by some integer multiple (say 16x) and then, for efficiency's sake, interpolating up/down the remainder using (Mitchell, Catmull, B-Spline, ...) is more common in practice.

 

i think mathematically it is the same, the LCR of 44100 and 88200 is 88200, so you can average by /2.

additional, "unneccesary" upsampling factors for the interpolation is of course also happening, and it a good compromise, but i am never sure if you can tell a best practice for that? it is interesting how different even the best apps are when it comes to the very last artefacts.

 

Adobe Audition is the best. run your file through there and you're g2g.