AudioSep

“I decided to try AudioSep: https://github.com/Audio-AGI/AudioSep on MultiSong Dataset.

I used prompt 'vocals'. I was sure it would be bad, but I didn't think it's so bad.

https://mvsep.com/quality_checker/entry/8408

 

I also tried it on the Guitar dataset - it's even worse - negative SDR. Maybe I'm doing something wrong. But I tried the example with cat from the demo page, and it worked the same as in there. So I think I have no errors.”

 

sdr: 0.33

si_sdr: -2.39

l1_freq: 17.62

log_wmse: 6.72

aura_stft: 3.66

aura_mrstft: 5.55

bleedless: 9.29

fullness: 16.58

 

Colab on GH probably gives unpickiling issue. You might be able to fix it be executing:

!pip install torch==2.5.0

After you execute all the installation-related cells.

Since then, probably something more about dependencies is also needed, like it ‘s coded now in the inference Colab.