Interesting!
wow! The output sample was really good, nearly native Persian speaker :)
I liked it very much.
By the way I'm in Iran too and because of challenges in gmail recently, I couldn't utilize it :(
Could you give me some guidance for how to download the pretrain model and use it?
I would be very thankful.
I'm glad you liked it! Thanks for the feedback.
There's a link to a Colab notebook in the README with all the necessary steps to get and use the model. I'll provide it here as well for easy access:
https://colab.research.google.com/drive/1e0BYsfMKRiikaA62umEyZejmivn0tTE8?usp=sharing
Also, there's a Hugging Face Space that lets you try out the model with no setup. Here it is:
https://huggingface.co/spaces/abreza/mana-tts
Thanks for your help. I tested the second link in Space for short and long sentences. It seems it doesnt work for long sentences :(
You're welcome! Yes, that's correct. The Tacotron2 architecture, by default, doesn't support long texts. To handle long sentences, you would need to implement code that splits the text and runs inference on the segments separately.
it is still great :)
Are you working on it to solve the long texts?
Thanks ^^ We have not officially released such a tool and are not planning to do so. However, I can share a code snippet I used for sentence tokenization: https://colab.research.google.com/drive/15MfqBg1ynJR6s6ixLXFdf9wTILkEMB31?usp=sharing