Sample rate of 40k, about 30 minutes of dataset (200 epochs might seem a bit low but @Blaise 's Titan pretrain allows for it). I made my original Toxicity era Serj Tankian model almost a year ago, back when we were still on RVC v1, and I wanted to re-make it with a cleaner dataset. Compared to the first version, this one is definitely better at rougher/distorted vocals, but without losing focus on softer vocals. I really hope you enjoy it!

🚀🔥

A collage of various inferences I did with this model, showcasing both clean and distorted vocals (some of the stems I used were not in the best quality, so if some parts sound eh it's probably because of that)

wow this is clean

Awesome model 😄👍

kinda sounds like Markiplier in some parts but I like that

maybe it's cause Serj sounds a bit like Mark anyway

titan magic

sure sounds like it

Seems to be a robust pretrain because my dataset is not flawless (but it is cleaner than the one I used in 2023)

The isolation models improved and my technique as well