[Discussion] Second stage model to make stems sound less "AI-like"

[rolaca11]

My main problem with the current spleeter outputs is that they have a lot of weird artifacts and inconsistencies. My idea is to have a second model for each output the current spleeter generates and train it with the original version.

In other words, train the second model to turn the weird result of the currently existing model into the original. Is this idea any usable? Unfortunately, I don't have any experience in tensorflow, or any other AI frameworks, so I can't try it on my own,

[junh1024]

You could:

• train a ML to add "stuff", but that would sum to more than the original.
• add production FX like distortion & reverberation, but that would also sum to more than the original.
• bleed each stem into each other, but that would reduce separation (this might be used for some commercial software)

There's also https://github.com/facebookresearch/demucs if you want a ML with different-sounding artefacts.