[UPDATED] Doraemon (RVC v2, rmvpe, 425 epochs, 24000 steps).

Trained till overtraining was detected with approximately 25 minutes of Japanese decent-quality dataset from Youtube.

Haven't tried other methods but rmvpe and mangio seems good.

Bear in mind that when the vocal/acapella's pitch is too high or low, the voice will break terribly, otherwise it should work fine.

You can combine both rmvpe and mangio audio outputs to get the best result, because in some places, mangio works better than rvmpe and vice versa.

• My recommended settings: (This is just my own experiments, you can go wild as you like)
  • Search feature ratio/Feature retrieval rate: 0.6-0.8.
  • Protect voiceless consonants and breath sounds: 0.2-0.4.
  • Hop Length (For Mangio): 128-512.