hello. i'm trying to use the legacy fakeyou tacotron2 synth and i get this error. i know it's because tensorflow 2 removed that function and i've tried installing tensorflow 1 by inserting !pip install tensorflow==1.15.2 in the cell but i still get the same error. here's the link to the notebook: https://colab.research.google.com/github/justinjohn0306/FakeYou-Tacotron2-Notebook/blob/main/Latest_Tacotron2_and_HiFi_GAN.ipynb

CC. @Justin

Works fine for me 🤷‍♂️

You don't have to do any of that in the first place

who told you to install TensorFlow 1 when it runs on TensorFlow 2 out of the box

i just clicked the link, put my tacotron id and clicked the run button

that's weird

you just said that you installed tensorflow 1 here

factory reset and try again

should work fine

okay so no active sessions

okay it worked now

thanks

i've trained the model with some extra letters but i did that last year and i forgot what to change to add the letters. here is the alphabet: AĂÂBCDEFGHIÎJKLMNOPQRSȘTȚUVWXYZaăâbcdefghiîjklmnopqrsștțuvwxyz the new letters are ĂÂÎȘȚăâîșț

you need to add embedding.weight into ignore layers list

okay i will try that

what file are you showing me ?

i'm synthetising, not training

okay let me see...

ooh are you training a base model?

if you'd like we could add it to uberduck officially and have your romanian base in the notebook

Is it okay if some sound effects sneak in the wavs? Most lines are clean but some of them have slightly audible sounds so I wanted to know if it might affect the overall quality of the voice and training. I already skipped some of the really loud sounds already (Training for pipeline).

It's my voice and I'm kinda afraid to share it :/

But if it works well with my voice, I will look into making a base model with public voice

understandable

i'd be glad to help if needed

I've already found a public domain website which has a guy reading poems for 24 hours

nice

Back then when I discovered that we didn't have openai whisper, but now we do so labeling won't be a problem

whisper performs kinda poorly on romanian though