Oh no!

her english bank has decent rs (not all of them but most of them are good) tho.. hmm, weird

I forgot which one it was but I think her Sakebi had good R's there.

the normal one should be better in tone but that's a whole lot of additional work

Well you can chain VBs together with subfolders and some slight presamp.ini modifications. I did it before and as I remember if you use the VCV versions It'll sound a bit more natural

Also if do have a good resamp and wavtool it might be better but that's usually a given for some

You can probably make Teto shift into maximun overgrowl too even ig :stare:

i made a custom arpasing teto for the job cuz arpasing is very straightforward

i just use TIPS and wavtool4vcv cuz tips uses PSOLA which i'd say is decent in shifting pitch in both directions

Oh yeah TIPS. It's a good one. Is it still in beta? Just tried looking and I can't find it

it's been released

it never got updated

attempt 2 of smash 64. he can say more than a word now but still big oof

I put your attempt through VoiceFixer and this is the result:

broiblebluaaaaaaaaaaaaaaaaaaaaaaaaa, it says

im starting to get good with voicefixer.

here have Microsoft Eva, the always-upset Windows 10 voice we all know and have a lovehate relationship with

and microsoft mark

Remember when scientists forced a worm into a computer? That was neat.

are you 80

No.

Just tried to train a model but apparently 12 entire files just decided to somehow be deleted. Fun 🙃

Sorry for the very late reply. I always put a punctuation mark at the end. Sometimes it doesn't give me the "reached max decoder steps" warning, sometimes it does. So I have to put two or more punctuation marks instead, right?

Is there a limit to how large batch sizes can be when it comes to training Tacotron 2 models? If so, what is the largest batch size that P100—or V100—GPU users are able to train TT2 models with in Google Colaboratory?

George Beard (Kevin Hart voice) is getting some stonks 😮

yeah (that or train the model for less time, if it's trained for too long it can go out of whack or trim the extra silence from the end if there is any or make a talknet model instead)

lmao i left colab training for real long and it asked me if i was still there and also gave me a captcha

Yup that happens

tried my native language on tacotron. i prob need a lot of samples for it to be good tho

depends on how much the spelling corresponds to the pronunciation, simplest case toki pona has near-perfect correspondence so works fine with 25 minutes