Is this with some sort of "base" voice dataset and then there's some application of another voice data set?

Sorry if I'm flurrying questions etc

nope

it's just one massive model

with all the datasets at once

as input it takes in the text and the desired speaker (and also emotions in the case of 15)

Right, which is based off of DeepMoji or smth...

I've just been skimming the related papers, so I'm just trying to get an idea if there's something that has global style tokens implemented aside from their interface

Like diving into an ocean, lol

the 4chan bronies' ngrok notebook (so called for the web proxy it uses) has both multiple speakers in one model and deepmoji

so if you wanna explore then

Right, so I got those links to work--pretty snazzy interface

I'm sort of wondering what additional effort would be required to add additional voices to the models that already exist

Or if it's actually straightforward (i.e. download the model directly and then additional training with a seperate ID or smth?)

i mean definitely a separate notebook for the separate architecture

and i think also train with all the characters that are already in, so it doesn't forget

we have some multispeaker models on the site. very interested in building features that let people collaborate on larger multispeaker models

I'd like to expose some latent space navigation features for those models as well

dataset collaboration is relatively simple, but I'm not quite sure what the best way is to do training collaboration

I sort of figured, I'll continue to dig around. Thanks for answering my questions

Good to know!

saving on an external server would allow for tag-team training, but that'd require coordination between 2 people in far enough apart timezones

hmm yeah

we can definitely do some training in uberduck, and maybe we could let someone spawn a training job with some subset of datasets

What’s the best multi speaker model right now?

If the models are just singular files, could you apply some sort of naming convention as well? Like you could use a standardized time zone + todays date. From there the interface could automatically choose the most current one respectively (or have the option to choose previous iterations)

I replied to the wrong comment

there's actually only one live on the site right now, kanye/jay-z/biggie

Oh really O.o what’s the architecture?

fastpitch