I̶ ̶m̶i̶g̶h̶t̶ ̶g̶o̶n̶n̶a̶ ̶o̶v̶e̶r̶t̶r̶ain if i̶t̶ ̶h̶a̶d̶ ̶g̶o̶o̶d̶ ̶q̶u̶a̶l̶i̶ty or ̶s̶ome̶thin̶g̶.̶ Made by me <@311764472886984706> Dataset made by me (8:57 Minutes) (2022-2023 songs from K-pop Band Kep1er includes Japanese Versions and includes 2 cover songs) Trained with crepe V2 24th RVC Model/9th RVC V2 Model. USE FEATURE RATIO TO 1.0 ALWAYS. Credit to me on Youtube & TikTok. Mangio Crepe 64 (350 Epochs): https://huggingface.co/CrimsonZockt/RVC-Model/resolve/main/Hueningbahiyyih.zip RMVPE (950 Epochs): https://huggingface.co/CrimsonZockt/RVC-Model/resolve/main/HueningbahiyyihRMVPE.zip And of course, The last retrain with Mangio Crepe Extraction. My model hit the overtrain after 350 epochs, I think so well. UPDATE: I retrain for RMVPE extraction, The new loss is 43. Who will proof with this?

Might gonna retrain for RVC V3 model.

Huggingface Link Up!

https://huggingface.co/CrimsonZockt/RVC-Model/resolve/main/Hueningbahiyyih.zip

You might gonna work for you.

I think i need to find the overtraining point of this.

Because i retrain with Mangio-Crepe.

The last one that i had from Original Crepe.

yeah. mangio-crepe is better than the rvc crepe. i recommend you to set hop length into 16 if your dataset has a clean, studio quality audio

i tried that in my kpop idol voice model and the result sounds like the real artist singing the song you like

I might go for 1000 epochs if i find the overtraining point.

I used Hop Length 64.

Couldn't do 16 for that

yeah, the modelname.pth with epoch has included step numbers so that you won't miss it out when finding the overtraining point

maybe the dataset has noisy or reverb like?

it's ok depending on your desired preferences

Aaaaaaaah! Forgot to set the save epochs to 50!

Overtraining Point

As a result, Same 350 Epochs & 10500 steps.

This means i retrain with Mangio-Crepe Extraction!

And that's it, All of my RVC V2 models retrain with Mangio-Crepe Extraction.

Time to make new models!

I recommended to use the Mangio-Crepe one of this not for Original Crepe train.

Demo Sample

I have to retrain for RMVPE extraction.

I think a new loss proves this one.

If 500 epochs is okay, It overtrains like brrrrrrr.

The new loss is 44.

I retrain for RMVPE extraction.

Huening Bahiyyih (Kep1er) (RVC V2) 350/500 Epochs

Oh well nevermind. 1000 Epochs incoming.

Maria Ozawa RVC V2 model is coming.

I think overtrain is completed?

Huening Bahiyyih (Kep1er) (RVC V2) 350/950 Epochs

I select 950 epochs.

For a lowest loss.

RMVPE Demo Sample

Let me know if Mangio-Crepe 64 or RMVPE better.

I think more models will coming up with Mangio-Crepe 64 hop and 300 epochs.

Bahiyyih çok güzel