Training a TTS model from scratch and Finetuning on private data

[VigneshBaskar]

Hi, I am trying to train a tts model from scratch on the LJ dataset following the instructions here: https://tts.readthedocs.io/en/latest/tutorial_for_nervous_beginners.html

I am new to TTS and May I kindly request someone to help me with the following questions please? I have the following questions:

1. It takes around 15 minutes to train one single epoch for Glow TTS on a single K80 GPU. Which means it could take around 10 days to train a TTS model from scratch on LJ speech data with a single GPU for 1000 epochs. Is this normal?
2. If I understood it correctly from the Humble FAQ: Would Tacotron 2 or Tacotron train much faster in comparison to GlowTTS?
3. Is finetuning the only fastest way the to train on a small private dataset of another speaker (~1 hour data)?
4. Given that my small private dataset is a male speaker, does it make sense to finetune on LJspeech dataset which is female?
5. Which one would produce better results: Finetuning on a private data of 1 hour or Training from scratch on private data of 5 hours?
6. Is more data better the performance by default or should I still fight with Hyperparameter tuning?

PS: I understand that the pretrained models are already available for LJ speech dataset. I just wanted to try it by myself :)
May I kindly request someone to help me with the above questions please? Thank you very much for your time

[erogol]

It takes around 15 minutes to train one single epoch for Glow TTS on a single K80 GPU. Which means it could take around 10 days to train a TTS model from scratch on LJ speech data with a single GPU for 1000 epochs. Is this normal?

Yes

If I understood it correctly from the Humble FAQ: Would Tacotron 2 or Tacotron train much faster in comparison to GlowTTS?

Mostly

Is finetuning the only fastest way the to train on a small private dataset of another speaker (~1 hour data)?

Yes

Given that my small private dataset is a male speaker, does it make sense to finetune on LJspeech dataset which is female?

Male model would work better

Which one would produce better results: Finetuning on a private data of 1 hour or Training from scratch on private data of 5 hours?

I'd try both

Is more data better the performance by default or should I still fight with Hyperparameter tuning?

More data > parameters

[VigneshBaskar]

Thank you very much @erogol