Present Arabic training results, not so good

Hello,
I make train tacotron2 + MelGan like the original notebook. I use public corpus 4 hours.
MelGan alone sound good, tacotron alone sound not so good and combination sound very bad. Can listen here.

wav-ar.zip (858.0 KB)

Is make modles + config + notebook here:
https://drive.google.com/drive/folders/1Gp355kY-xA3AVH38yXmAWamV0FRKc0RX?usp=sharing

Appreciate help very much

what is the dataset?

also you need to give permission for your notebook.

Is make share here:
https://drive.google.com/file/d/1LSJmMZux36_gDyDIGk7jOieUc88Phunh/view?usp=sharing

DB is get here: http://en.arabicspeechcorpus.com/

still it asks for permission :frowning:

Please make try now

https://drive.google.com/file/d/1LSJmMZux36_gDyDIGk7jOieUc88Phunh/view?usp=sharing

your notebook fetched data from your GDrive and it is not there for us. Therefore cannot see the arabic model.

I make shared here:
https://drive.google.com/drive/folders/1Gp355kY-xA3AVH38yXmAWamV0FRKc0RX?usp=sharing