Description
Made by me The training data was curated from Joji's post-Pink Guy songs, capturing his modern essence Tensorboard was used to ensure the selection of the model with the minimal loss during the training process Settings: Crepe Hop Length: 16 Batch Size: 8 Epochs: 1050 Dataset Length: 9m 16s <
Comments

Sample: d4vd - Romantic Homicide

Model is at 25K (red circle)

bruh u overtrained by a lot

unless you got the model back to 25k and saved that

then u fine

but if not well then the model could've been better

v2 also trains quite fast sooooo..

Sounds good to me tho. I've been following the recommendations to stop once loss/g/total starts rising

yeah but it also has like a 4k step difference since it statred to rise

that's like 4 epochs

i think

but yeah it sounds good

I'll try from a previous epoch and see if it sounds any better

when it reaches 25 k stop it

just for u to get the best quality

or try increasing the dataset

I'll make a comparison between 21K and 25K, good test?

is this model now 25?

Yeah it's 25K

ohh then u good

that's the best it can go

That's where the red circle in the img is to show where the model is

ohh yeah you are tottaly fine

Ohh ok thanks!

thats the best quality u go

got

np
Add a comment
Samples
1. Singing
Male
English
2. Singing
Female
English
3. Singing (Dry)
Female
English
4. Singing (High)
Female
English
5. Singing 2
Male
English
6. Singing (Dry)
Male
English
7. Singing (Dry, High)
Male
English
Pitch
Users also tried
More to explore
Loading more