the exact number depends on the length of the wavs but i think 20~30 is about the max range

try duplicating one of the lines in the list.txt so it's an even 270, then training with batch size 30 if that doesn't work then i guess 15

and again, arpabet

Noted. I'll try it. Thanks for the help.

it ran for 3 seconds and gave up lmao

oh? please send the error

I'm running it with batch size 15 now, which is going as expected, but (with batch 30) it essentially just stopped after the initial loading of the model. Progress obviously didn't make it past 0%, and it didn't actually start training an epoch.

...interesting

It's suddenly gone from constantly saving the worst iterations to saving halfway decent ones. I hope this is a good sign...

Hi, hope someone can guide me here, I produced models using both TacoTron 2 and Talknet. They both turned out pretty good except the pace and some tone issues. For the talknet using the controllable synth notebook here. Assumed the reference audio could fix my issues, but when I use reference file of same speaker with desired style pace all I get back is gibberish. Using 6 second reference. Isn’t that what reference audio for? Am I using wrong? Does TacoTron 2 have way to match style ? Thanks for any help.

Booted after 294 epochs 😂 Still sounds really robotic, but pronunciation might be improving a little? Thinking I might give arpabet a try, and try batch size 30 again. Any way to get it to start sounding less robotic?

It seems like it depends on how good the audio quality is, and maybe how long you train it? I did a practice run with ~1 1/2 mins of decent quality audio of my favorite singer, and trained it to 250 epochs. It sounds very robotic, but perfectly matches his accent and speaking style. Maybe you need more audio of the style you want, and put that in with the what you already have, or maybe you'd get better results from using only the stuff you want it to copy. That's my best guess, anyway.

Nevermind, it still falls apart on anything remotely complex, like this series of short sentences. If I leave those periods as commas, it just turns into unintelligible noises. 🙃

for tacotron its all about val loss

if you get a good val loss (usually sub-0.95 or so) and your model still strokes then a clip isn’t aligned properly

(0.095 you mean)

By aligned do you mean a clip that has a bad transcription ?

it can be a number of reasons

Hi, so I don't think that any model from uberduck will work with grid3, I know you're positive it might and you want to try, but I honestly don't think it will. It is possible to download Talknet or tacotron2 (uberduck uses this), and then use that as a text to speech.

You do need a decently good graphics card though

could someone send me the link to grid3?

Sure, when you try it out, you'll see why I suspect it won't work

I see

Yeah, sorry @Komori-uta Kitsune , you could still try to download and setup either Talknet or tacotron2, if you do I can give you a voice (of wolfychu) but her voice is very high pitched and you might dislike it, but zwf might give you the model you want

Could also be very risky, as I've not gotten permission of them to make it.

But that goes for almost any model

how many wav files did you use? And did you change any of the talknet training settings? Do you remember how long you trained for? Because the result I'm getting from talknet seems worse

I just redid the training for Aerith. 600 epochs and 6700 steps for hifigan. Index.wav = with reference voice. with_ref.wav = with reference voice. aerith3.mp3 = my voice reference