Description

Trained on ~2m8s of data using rmvpe. Demucs 6s isolated most of the vocals, but if there's something that can isolate violin there would be more usable data. Transpose -12 and low volume envelope mask (~0.2) recommended. The feature ratio determines the amount of distortion. At ~0.1 it just sounds robotic, but it becomes very distorted at 0.8 or higher. Please credit me if you use any of my models.