Clean Inverted Vocals

Both of these models are specifically created for cleaning vocals (acapellas) from artifacts after phase inversion - a process where the official instrumental is subtracted from the original track. Standard models for vocal extraction usually cannot handle the specific distortions left behind after such a procedure (especially if the sources were compressed, for example, in MP3).

• bsroformer_cleaninvert (by gilliaan): A newer model (a fine-tuned version of "Revive"). It is tailored to remove "pixelated" and smeared artifacts of MP3 compression, as well as crackling and clipping. It makes the inverted acapella as close to a clean studio source as possible. The author recommends first processing the vocals with the "big beta 5e" model, and then cleaning up the residuals with this model.
  Original weights

• Mel-Roformer invert-clean (by becruily): An earlier model for a similar task. It perfectly handles clicks, sharp peaks, and "stubborn" instrumental residuals. It focuses less on the fullness of the vocal sound and might slightly muddy the audio in very loud sections, but effectively cleans up the junk.
  Original weights

Although both models can technically extract vocals from regular (non-inverted) songs as well, standard models are better suited for these purposes. The main specialization of these two neural networks is precisely the "deep cleaning" of complex artifacts left behind after subtracting the instrumental.