@User Hey did you work with Hifi Gan? before ?

I haven't worked with it much but I'm using a pre-trained hifi gan on uberduck.ai.

what will we hit 400?

Oh I see so effectively you just pass on your mels to hifigan

I meant people online.

Yes, I'm using it excatly as I used waveglow

and like you don't really need to train the model e2e with hifigan

I found it's much higher quality than waveglow.

i found a colab waveglow tacotron2 but its only the LJSpeech

it would definitely improve the quality to train e2e but it's not necessary to have satisfactory output in my opinion.

@User okay I guess ill train on fastpitch and see what happens, some of the data processing is annoying is there a script to get it into a form that is similar to LJSpeech

LOL, that Tom Scott video is exactly what the interactive part of Uberduck is. GPT3 into a synthesizer.

Yeah, the data processing is extremely annoying. It's the most time-consuming part.

Do you have any transcriptions now?

Also btw is anyone into anime culture here?

do you have screenshots how you work on tacotron2?

I have a tool that I've written for myself, and I know there are some other colabs.

Yeah I have Scarlett Jo and also some anime voice actress

just gotta get around to processing and cleaning some data

if i clone someones voice in colab but sounds like dying person

and starts speak gibberish

This often happens if you don't have enough data.

it needs 10 or 5 secs

ah, I see, the corentinJ one.

yeah that one works differently, small amounts of data are ok.

But it's not that great for voices that it was not trained on.

@User I heard he nurfed that code so that it was hard to reproduce results

The thing I'm working on now will hopefully provide similar functionality.

wich voice you working on?

I am trying to replicate the results in this paper https://arxiv.org/abs/1910.10838