@hecko The two images you were asking for?

ah so

you should keep

datasets

empty

because that expects one voice in each directory

sorry for making it so confusing

Going back to the subchannel

Man it’s been so long why has no one accepted to do Lalala voices for me I ask a few months ago and still up to this date still no one has done it for me

no one is obliged to do any voice suggestions, they are simply just suggestions

Seems you are the only one, who ever will make this voice

Ya but it’s taking forever and I am being so lazy about it.

Voice suggestions =/= guaranteed

In that case you'll wait forever probably 😄

At least providing dataset will make it waaaay easier for those, who train

I'm not sure if I should put this here or in support, but I'm working on updating a model I made and the dataset is literally getting so fucking gigantic. like the first one I made had abt 460 wavs, and I'm not even done collecting data and I've worked my way up to like, well past 1200. so yeah basically I just wanted to know if anyone else has trained an ai with over 1k wavs in its dataset or if it's going to be absolute hell to train this thing lol

both

i think tom hiddleston was trained on 1000+ wavs

several but they're all multispeaker so idk if they're inherently sucky or if i just haven't trained them for long enough

though really it's like 100 epochs on 100 wavs is the same amount of training as 10 epochs on 1000 wavs so you can train for about as long and get about as good a result the difference being that with more wavs it can reach higher quality, and that it'll save 10 times less often (so use a notebook that lets you change that, either arpabet or pipeline)

okay cool ty!

Hey guys, This is an idea that has been kicking around in my head for a while. Has there ever been any discussion about making a channel to share data sets? I'm actually surprised people don't do this by default, considering that it would help new users to discover what makes a good data set so they can train their own model, and it would also help to experiment with different models on an already created set of samples.

If you can't see it, you need to get the @Agreed role in #842148452464853002

unless I'm missing something, I don't see a data set channel here.

It's under the

COMMUNITY

tab

Ah ok, found it. Had to react to gain access, so I missed it the first time lol.

Sheesh

REAL

Internet moves fast

So what I can figure out, it's entirely T5 based and doesn't use a prior?