the main tacotron network outputs a spectrogram (timeline of sound frequencies), and by default uberduck uses a preset hifi-gan model to turn it into audio however since that hifi-gan was tuned on a different voice it can sound artificial (e.g. my model of me would go super low at the end of a sentence) so you can instead tune a hifi-gan on your own audio and submit it (via ticket) and it'll sound that extra bit more realistic but i'd only recommend doing this if your tacotron model already sounds pretty darn good

@Echoes @Desuka can i have your akane datasets? I'm making a big akane model and i think it would be cool if i could add in what you guys have too

which type of model is best

tacotron, etc

getting back into this stuff again

talknet if you can give it exactly what it wants

tacotron if absolutely anything else

sure! want me to put it in #835647732453605376 or in DMs?

either one works, but if you put it in #835647732453605376 i'd like you to ping me since i have that channel muted lol

Where’s the ARPAbet colab?

The Tacotron training colab with ARPAbet built in

Let me check my drive

Oh wait, found it

is there a tacotron 2 hifigan training note book?

hopefully thats not a stupid question lol

You can fine tune hifigan after training The notebook is in the notebooks channel

ok which one is the training one

im like

brainded when it comes to this stuff

its listed under hifigan then training

🆘

Optimus Prime V4 at 50 epochs

What does Descript use for their Overdub voices, I tried one myself and they sound so realistic, do they just have a better way of synthesizing voices or something?

it's proprietary

they won't tell us what it is

but part of it might be that they're synthesizing at 44khz vs our 22khz

and also not using tech from 4 years ago

vits can't come soon enough

can the uberduck pipeline be used to train a 44khz model?