add a basic trainer and dataset

[manmay-nakhashi]

• huggingface dataset
• basic trainer

[manmay-nakhashi]

to test or train just run
by default it's using MushanW/GLOBE dataset 23,519 speakers and covers 164 accents

python3 train_e2.py

Note: need to add a path to vocab.json

[lucidrains]

@manmay-nakhashi Manmay! i remember you now from the natural speech work we did together some time ago

thanks for the PR! I will check it out tomorrow morning 😄

[lucidrains]

@manmay-nakhashi hey, looks good! 😄 do you want to try pulling and integrating the text as well?

[manmay-nakhashi]

Sure I'll do that.

[manmay-nakhashi]

[manmay-nakhashi]

@lucidrains it's ready

[manmay-nakhashi]

i'll write a inference script next so we can do some quick experiments.

[lucidrains]

nice! it looks good, but in the paper, they didn't use a tokenizer and just went character level

i was thinking we could just use utf character ids? (could remove the tokenizer and vocab.json altogether) keep it simple

[manmay-nakhashi]

@lucidrains changes are done

[manmay-nakhashi]

@lucidrains resolved all the suggestions

[lucidrains]

@manmay-nakhashi thank you Manmay!