Yulia Volkova (t.A.T.u.)

Description

Voice model of Yulia Volkova, trained in Colab to 500 epochs, on a dataset of processed studio recordings, speaking, and live, focused on getting her to scream! In higher ranges, at index ratio 1, it seems to be working! This model is a rework of my old models from so-vits and in higher ranges may even be superseded by jpopkaraoke's, for which I (very happily!) provided the dataset. I will be uploading a 500 epoch Lena model as a complement to their 300 epoch model. Using one or both of these Yulia models may be beneficial for different vocal ranges. jpopkaraoke's Yulia/Julia model: Notes, there are some audio that neither of these voice models will be good for converting harsh vocals like Chester Bennington screams etc. Inputting vocals that are screaming can make the pitch dip down instead of interpreting the correct pitch. This is because of the way pitch estimation works in RVC and I haven't found one model or conversion method to combat it. Besides that, there is a delicate balance between getting her to scream and not, and playing with index value 0 and 1 respectively seems to do some kind of magic. Multiple takes may be needed with different index values or even re-singing troublesome parts of acapellas yourself. For my own covers, I re-sing everything multiple times to avoid this. Trying a wide range of index values will give interesting results. Higher index values tend to sound better upon testing, like studio recording, however again, like all other models, your mileage may vary. If you want a screaming model, you may have to try a dataset of harsher vocals! I am willing to do this if anyone asks! I may need to rely on donations at this point to increase my colab subscription though.