I trained this to 200 epochs a couple hours ago. It probably still needs more data, along with more breath removal.

I'm torn on cutting out drawn-out yesses from the dataset. They're kind of a signature trait for this character.

maybe type them in a way that the dictionary won't pick it up

if you're using the arpabet notebook

let's go voice cloning WOOOO! transcript: This is a test of {T AA1 K OW0 T R AA2 N} even though I've used it multiple times. This model has thirty minutes of data to train on using the {EH1 L} {JH EY1} speech {T R AE1 N S K R IH2 P T}. {Y EH1 S}, this is a voice clone. I'd say it's {P R IH1 T IY0} spot on.

Voice cloning, nice!

Now this looking spicy NGL lol. Nice job

quite epic

Have you tried talknet? apparently you can even use the HiFigan of tacotron to influence the quality of your talknet

It’s also similarly set up the same way tacotron 2 is

Sounds cool though

Hello

so another 20 minute test. This time of my cousin. Sounds perfect

i got her voice because I think I found the perfect way to troll online

hontou desu desu de

^ that was the result of training 15 files of japanese speech to 275 epochs (resulted in validation loss of exactly 0.1 for some reason lol)

I have tried talknet before but I might try it on singing samples instead

this is one I made on singing samples I recorded

exactly why I wanna do it with singing !

you can do it!

Is this singing or not singing btw

BC it sounds pretty good

Nvm it is singing

@HolyArapaima consider importing it to nnsvs that'd be cool

a 21 second long dataset looking too suspiciously normal!!!

What voice are you doing Matty?

a fnaf character, again

Which one

pigpatch from UCN

litterally 5 wavs