things like that

is that still 'usable'?

transcribing an audio is very boring

i think something went very very wrong

also, does colab have restrictions on upload speeds? I typically upload at ~20Mbps, but in colab i barely go above 7Mbps, which turns my 50MB dataset from being a 20s upload to a 70s upload

@_n you disabled pronunciation dictionary?

m e m e

i think it might have just been a problem with my model. i trained and synthesised using the simplfied notebooks, and i was getting good graphs but no audio

im going to train the model again, this time using the original training notebooks

Did your GPU have K80?

nope, so far i've only been getting T4s

Ah, okay

ok yeah, i think it's actually working this time, because last time my epoch 0 graph was pretty much a straight line

yes

why do you want to finetune hifigan anyway?

always that the best you can get

i'm not sure what the context here is

maybe you should read the code for generating the alignment graph to understand why that happens

I feel like I'm the only one writing here who gets the concept of the alignment graph

and there is non determinism in generation of the alignment graph

How can i make voices in uberduck?

Can somebody caption this?

there's a few good ways to answer your question! #841422801965416538 is a wealth of information and in the pins of this channel there's a wonderful tutorial ( that you can view here )

alright, got some halfway decent results

not great, very grainy, but you can mostly tell it sounds like wheatley

i was playing around with the sigma and denoise strength, very low sigma makes very tinny sound, and very high sigma makes very crackly sound, somewhere between 0.7-0.9 seems to be the sweet spot between crackly and tinny

denoise_strength seems to not have any noticeable effect, regardless of the sigma value

using raw_input = False or any kind of ARPAbet strings seems to completely break the synthesizer

it says "This is the sound of wheatley at ninety something" perfectly, then just turns to gibberish

you'd need to train specifically on the arpabet input