search

p0p4k / vits2_pytorch

unofficial vits2-TTS implementation in pytorch
https://arxiv.org/abs/2307.16430
MIT License
465 stars 81 forks source link

train files for vctk 0.92 #48

Closed AWAS666 closed 10 months ago

AWAS666 commented 10 months ago

changes I made to the original file list to be compatible with the newest vctk 0.92 version, not sure if those should be kept seperate or merged with the original ones.

tldr: I made a little script that deleted any entries which have no audio files

I8dNLo commented 10 months ago

You also can provide updated config to run this on actual vctk dataset

AWAS666 commented 10 months ago

You also can provide updated config to run this on actual vctk dataset

oh it's just adding in the new file paths, depending on whet ever these should be the main files or not it can be added, otherwise just rename them to replace the current

p0p4k commented 10 months ago

Thanks for the updated filelist!

AWAS666 commented 10 months ago

Thanks for the updated filelist!

should we update the original files, since these are currently new ones?

p0p4k commented 10 months ago

We can keep both versions and add a new config file for vctk_new ? Thoughts?

p0p4k commented 10 months ago

What if we add a function in dataloader that skips missing files and stores them in a text file under the logs folder of the training for the user to later verify that? If number of files available is less than ~30, we can assert error.

I8dNLo commented 10 months ago

It could be just a warning to say like 'Looks like you are using old version of dataset'

p0p4k commented 10 months ago

It could be just a warning to say like 'Looks like you are using old version of dataset'

Good practice is to generalize your code for future versions. But for now a simple warning like this could work too.

AWAS666 commented 10 months ago

I mean the current is perfectly fine for anyone using their own dataset, it's just about the vctk versioning. I'd do just the new config for now and maybe add a flag to ignore the error altogether?

p0p4k commented 10 months ago

Cool. Did anyone train back yet? How are the results?

AWAS666 commented 10 months ago

I've reset after some first trial and error runs to see what dataset works. Currently sitting at 124 epochs and the sample in tensorboard sounds decent.

Though I cant do any inference as I've moved to windows and I got no espeak there :(

p0p4k commented 10 months ago

I've reset after some first trial and error runs to see what dataset works. Currently sitting at 124 epochs and the sample in tensorboard sounds decent.

Though I cant do any inference as I've moved to windows and I got no espeak there :(

Thanks for your feedback! Good to see something positive about multispeaker training. You can always infer on Google Collab for free, temporarily.