Is like the max size a model can be 322mb?

Idk but that is normal

got the buddy lines, can you cut these up? these are a LOT!

who's Buddy from

Kick the Buddy

But like that’s what makes him Rick

sadly that's what you need to do

what da machine learning

Dang

@User quacked: Number 14. I don't learn to shut my god damn mouth.

@User I appreciate that you tried to improve our turret model, however when you did so you forgot to fix voice files that cut out. We originally had this problem too, we had to use the other conversion method instead of the one you used because this happened.

Under our version, the entire line was able to be emitted.

why's zorak coming out like this

Thank god it didn't come out horrible again

Pete's not looking good either

Lookin Awesome

so how do dashes work between words? what about symbols in general?

training moo moo the magician right now

i've had so much fun playing around with uberduck, this stuff is bananas! thank you for all the work you guys do 🙂 i have a question for anyone who's had experience training and modelling singers/rappers i'm out of my depth here, so i hope the question makes sense i take it that the more natural sounding voices are ones that have larger datasets and longer training time. where are you getting these from? are they artists' released acappellas? are the chances of creating a natural sounding voice determined by the amount of acappellas released of any artist? would using vocal extraction software aid the process? have you already been trying that, using something like Spleeter? if you have, how has your experience been? thanks in advance to anyone who takes the time to answer these questions and please let me know if i may be fundamentally misunderstanding this stuff

thank you pi 🙂 i will give this a read

To give you the teal deer also, you are correct! The chances of creating a natural sounding voice are indeed determined by the quality of the audio used I don't know where people get these studio acapellas from but what I do know is that for voices with unusual quality to them ( i.e shouty voices and voices with a lot of variation in the tone ) the more you can collect of that kind of clean audio the better it's going to turn out. Vocal extraction can work but it's no studio sample that's for sure! ( Though, as a footnote, a data set made from extracted audio can work! I have a model that utilizes quite a few clips from a cartoon and he performs nicely )

awesome, thank you for expanding on that!

do you have the cartoon voice model shared on uberduck for me to take a listen??

I'm starting to think I should plan to make Marty and George McFly, especially Doc Brown from Back to the Future as TTS Voices, but I don't think if I could gather 5.1 dialogues from all 3 movies since I have 53.2 GB left on my computer. Each of those movies have 3 gigabytes, and all together can be 9 gigabytes. So, after I downloaded them would be 44.2 GB left on my computer. And that's gonna suck because if I put them in my RX 8 Audio Editor, it's gonna waste more memory. So, I'll just let people know if they could gather 5.1 dialogues from Back to the Future movies. I'll DM you the links to BTTF movies that has 5.1 audio of course.

I do not have him shared publicly at this time but I do have a few samples I generated with the Hi-fi Gan notebook!

length of the lines in the transcript also makes a difference! this model, for example, tends to do better overall with smaller text in the form of 1-3 words because the characters I trained him on also had a lot of tiny lines

How does Tractron like reverb? I’m pulling vocals and there’s a lot of layering and reverb on them.

clips with reverb makes tacotron cringe