send me your transcript

Question: Why are TalkNet model users trying to use the wrong notebook? I updated #841437191073955920 a couple days ago.

Maybe I should remove the link to the CPU version of the synthesis notebook.

Maybe archive the ones that don't work for now

I put the link to the CPU notebook in another message

Thank you for the help and I have another question, is there any limit on how many .wavs you can use to train a model?

in my experience with colab using an ~1-20s bell curve distribution, the colabs are too burdensome to use >15k wavs. Ive recently upgraded to the pro-est pro so they're more forgiving but otherwise I would suggest staying under that and I'm trying to find the sweet spot for how under that currently

Oh, thank you for explaining! I've been thinking of about like 1200 .wavs! And also maybe stopping the Hifigen at 10.000 steps

im toyed with a lot of parameters, most recently taking the spectrogram one way wayyyyy out (400-2000 epoch spread) and didn't notice a significant degree for it. hifi by numbers seems super erratic to me so I haven't gotten a feel for when it's "done" do let me know if you notice something special with it

I will definetely! Yesterday I did it for the first time to train a Britney Spears AI, my mistake was that I used one song and made about 40 wavs and stopped at 2200 and It sounds all sorts of wonky - I will redo it later today with what I've said above

im working on rupaul now using a 'truncated' 1k wavs to see what quality I can get comparatively

while I understand overtraining is a thing it seems harder to argue against it 0given the context of voice synthesis so i've been taking the default duration and pitch epochs x10 (200 and 500 I think respectively) obviously improvements are increasingly negligible but ideally we do want them to sound literally the same as the wavs theyre trained on? 😅

Yeah, we really do 😂 I really want mine to stop sounding like there is decades of smoking damage on the voice so I'll definetely use more .wavs, turn the epoch up on that one step (I am very uneducated in these things and can only know what I mean when I am on the Colab page lol) and definetely let the Hifigen do It's thing longer

imma try extending out the hifi next too - lets compare notes

as an example here are various donald trump models where I was playing with the spectrogram parameter... (to me) it doesnt sound significantly better given the effort involved. and then wmetallicreduction being the toggle on the controllable talknet notebook

They all sound pretty impressive! I'll send the files as soon as I am home, but I never even got any sound for my model without wmetallicreduction being on. It just made very unpleasent noises without that.

ok. saved a model at whatever it is and 4500hifi and then ill report back after its run a significant time

Ugh.

Can someone help with this

You should talk to @Justin about this problem

im not technically smart enough yet to understand whats going on under the hood for the metallic noise reduction on the controllable talknet but as of now my monies on that being the best area to seek further improvements

Hi, I'm currently trying to use this notebook. But I get this error when I try to run the interface. Anyone encounter anything similar?

Just a guess but make sure the runtime mode is set to gpu

Well this is how it is currently. Is this what you mean?

Yep

A100s don't work yet

aim for a P100 or V100

T4 aswell (duh...)

Wait why