How to train a Tibetan-TTS model about pytorch-dc-tts

[tugstugi]

You have to create a Tibetan dataset which should have same format as the LJSpeech: https://keithito.com/LJ-Speech-Dataset/

It means you need many audio snippets up to 10 seconds and also its transcriptions. For Mongolian, there wasn't any dataset. So I had to create one (MBSpeech) from the Mongolian Bible audio book. For manually splitting 5 hours from the Bible audio book and transcribing it, took me around 30 hours.

If you have created a such dataset in LJSpeech format, you can also use it for all other open source TTS projects such as https://github.com/Rayhane-mamah/Tacotron-2 and https://github.com/NVIDIA/tacotron2

They will have also better audio quality than this project. The only advantage of pytorch-dc-tts is, it is fully convolutional and so really fast to get results. Change both parameters text2mel_max_iteration and ssrn_max_iteration to 50.000 or even lower, then you will get feedback in few hours whether your dataset is working or not.