nice, good luck. it's a lot of fun

@User Here's the glowTTS repo, They have a notebook ("Inference.ipynb") that you can use

I've never tried it myself

How to run inference.ipynb

Can we use pretrained model?

yes, they link a pretrained model in the repo README.md

Oops, I forgot the link

Can we use the michael rosen datasets?

You're asking ZWF to hand over his datasets for you?

i made a michael R dataset already

you put together the audio, but you didn't transcribe it, which is the time-consuming part

how to transcribe it?

you can use Descript or you can try to do it yourself

I've found Descript to be really useful

what is this?

Although you still need to go back through and correct the transcriptions

It's a program that lets you edit audio like text https://www.descript.com/

So yeah, basically the input to the model is a text file that looks like:

path/to/wav/1.wav|Transcription of the first file.
path/to/wav/2.wav|Transcription of the second file.

oh

where each individual wav is between 1 and 10 seconds

also we can write

I make my datasets using Descript, so if you create a Descript project where each paragraph contains 1 to 10 seconds of audio then I can easily make the training set

michael rosen needs to be updated vo.codes

yeah, sounds like high-fidelity models are coming on vo.codes though

cuz they have more funding now

@zwf have you used Nvidia Nemo?

I've seen it, but never used it

Going to try it out, been downloading the docker container for the last hour 😅

nice, let us know how it goes!