New LibriTTS-R voice could replace the current LibriTTS voice

[StoryHack]

A dataset has recently been released which took all the recordings in the LibriTTS dataset and used a restoration model to remove background hiss / echoes / etc. It's called LibriTTS-R. A little bit of work would need to be done to prep this new dataset, but a new high quality voice could be trained from these improved recordings. They are available at http://www.openslr.org/141/

However, I'm guessing it'd take quite a bit of GPU time to train such a beast. my piddly RTX 3060 would take forever.

I am curious how long the existing libritts voice took to train and on what hardware.

Also, does anybody have a good colab notebook for training? I wonder if it might be affordable to pay for the processing units to train a new voice in a reasonable time. I've just never written my own notebook.

[rmcpantoja]

HI @StoryHack,
See the notebooks folder in this repo.

[StoryHack]

lol, I'm an idiot. Thank you.

[StoryHack]

I see elsewhere that the LibriTTS-R dataset is in fact being trained.

[synesthesiam]

Currently on epoch 254 🙂