Damn?? When i tried to install coqui on colab it gave a numpy error iirc

can you show me the steps you followed? I know just enough about Python to get shit working, and this really wasn't super hard surprisingly. I just now figured out how to list all the various models it comes with, basically prebuilt voices, and download some to try out. Next step is actually training.

TBH I wonder if Colab is really the best option for training these models moving forward. I get that not everyone has access to a good GPU, but from what I understand training can be accomplished on CPU. I figure if you're willing to wait a few days for the training to complete, especially with a pre-trained model, you would get a much more stable experience. I myself do not have access to a good GPU, but I want to try this anyway to see how long it ends up taking.

ah, i never really considered training on cpu before i'm not exactly sure how to go about it tbh i think it would be beneficial, though, since colab free has strict limits, and colab pro doesn't let me use my debit card, or paypal

or a prepaid credit card

The great thing about Coqui is that it seems to automatically detect whether you have a GPU or not, and configure itself accordingly. Training literally involves creating a configuration file in .json format, and pointing the training script to the path of that file. It takes care of the rest. All of the models behave the same way, so you don't need to restructure steps to train a different model.

What OS are you using? I can help you get Coqui set up.

i'm on windows (and linux and macos, but primarily windows) ig i can follow the windows tutorial there?

that sounds like incredibly easy training though, way easier than espnet or even default tacotron looks like i could train vits with it too? and several languages, including french and japanese

yeah, I've just started training a model. I'm really new to doing local training, I have mainly only done it on Colab thus far, but I'm really surprised at just how easy it is. I don't think my model is going to turn out that great, I only am training for like 200 steps, but I just wanted to see if it would even work at all. Keep in mind that this is on CPU, so it will be slower, but my training is moving along at a pretty good pace, and this is without a pre-trained model.

Nice!

It'll probably be faster if i get a gpu (i planned to for several reasons)

so when you're writing your transcript should you write numbers as symbols or spell them out? or does it like not make a difference

the training notebook will convert them to words but personally i like to write them out

okay cool thanks

redid my japanese tacotron notebook and updated the tutorial https://docs.google.com/document/d/1PXoczX2oq2g24_CMlTiSLb5s3CUh2b7QVlrA7CViNPU/edit?usp=sharing https://colab.research.google.com/drive/12Ke1vHDpNPGSSfFTEUpt8t28uJjN6ZDk?usp=sharing

OOoh a Japanese one!

yeah! it works pretty well

Maybe I can finish with Serial Experiments Lain

At least get Lain done

incheresting

Lookinng at part 2, what format does

wavs/<name_of_file>.wav|[Japanese text here

mean?

Hiragana? Romaji?

would be nice to get jp elirin properly transcribed but i cant japanese for shit

japanese text in kana or kanji

i should specify that, one sec

if i provided the dataset could someone do so 🥺

how long is it?

Ah, I'm not that familar with those. Haven't doen the transcript for it yet so that won't be painful of redoing it