Both are really good

is the talknet model also on the website?

also have you seen rad-tts

Yeah, when you click on the music icon with Weeknd selected it should become the Talknet model.

Also I'm somewhat familiar with it but haven't made a model on rad

can also access it through this link https://uberduck.ai/#voice=weeknd-tn&mode=tts-reference

example with the talknet model

i used this audio as a reference

lyrics are from this paper project page https://jerrygood0703.github.io/KaraSinger/

oh I just realized I tried it without first splitting the vocals, one sec

You know this has been on the site since I put the tacotron model on the site right?

that test audio sounds pretty awful, thing with talknet is it needs really good ref audio or it turns out like garbage every time no matter how high quality the model is. A lot of the ref audios on the site aren't great quality and reflect that!

I tried it again after splitting the vocals

do you mean this one sounds awful for the test audio?

Yeah that's still pretty bad, the pronunciation is almost not there at all, words are way too shoved together.

I found the Steve Ballmer clip posted in #859823464847638538 is a good test.

it uses the full pitch of the model, doesn't get confused about the pitch of the ref audio.

oh ok

creepy

Opinions?

pretty nice

talknet notebook upload reference not working

are you using cpu or gpu synthesis

cpu is very picky on its length

<10 seconds for cpu

otherwise it'll fail

cpu

ill check again

ohhh that'd explain why it refused to work back then

I just finished training a dataset on Tacotron and I don't think it saved the .tar file