Description

This model was trained using RVC v2 (RVMPE 300 Epochs) after building the dataset with Koriass most recent albums (MP3 320kbps) quality. Each tracks were isolated (Vocals/Instrumental) using UVR5 (BS Reformer ViperX 1297/1296 & MDXC23 Karafan Ensemble method). It's one of my first trained models using RVC v2 (RVMPE). Not every tracks were used because of artifacts and back vocals that were too much present. Even models like HP2-Karokee and BVE model will almost always leave some residue that you don't want to have in your dataset. For additional details for you nerds like me out there each and every vocals were individually isolated and cleaned (De-Reverbed with UVR-DeReverb-DeEcho, then De-Noised manually with iZotope RX 10 Spectral Denoiser & Supertone 'GOYO' Voice Separator). Model trained using Pretrained 'F0Ov2Super40kG.pth' and 'F0Ov2Super40kD.pth', RVMPE, 300 Epochs. Unfortunately I did not have the proper skills to evaluate the tensorflow at this time, it may be possible that it was overtrained, however the dataset was 1 hour long.