Dataset licence

[neonbjb]

Dataset licence

[neonbjb]

The dataset consists of thousands of audiobooks and podcasts that were scraped from the web. Many are copywritten, which is why I am not releasing the dataset.

If you know or believe the laws in your jurisdiction will consider ML models as extensions of their datasets, then you should consider Tortoise license encumbered and you should not use it for commercial purposes.

[C00reNUT]

Thank you for the clarification.

Just one more thing, I am asking because I would like to use the train_ voices

This repo comes with several pre-packaged voices. Voices prepended with "train_" came from the training set and perform far better than the others. If your goal is high quality speech, I recommend you pick one of them. If you want to see what Tortoise can do for zero-shot mimicing, take a look at the others.

I just want to be sure that they are not 'exact' 1:1 copy of the original voice, because maybe the generalization of the model could be fine according to the law, but I wouldn't be so sure with the exact voice match

[neonbjb]

This is a good point. You should not use any of the pre-packaged voices for business purposes for the time being. I will re-open t his and investigate which voices have copywrites attached to them and remove them.

[neonbjb]

FYI: LibriTTS and HiFiTTS datasets were used to train Tortoise. If you are looking for license-free voices that will work very well with this program, use one of those.

[C00reNUT]

FYI: LibriTTS and HiFiTTS datasets were used to train Tortoise. If you are looking for license-free voices that will work very well with this program, use one of those.

Excellent, that is a very valuable information. There shall be plenty of public domain options, it will be just a bit of hit or miss trials

[Aspie96]

Just as a (probably) dumb (related) question: is there any reason to favour those datasets over LibriSpeech or some other dataset based on LibriVox (maybe a public domain one, since LibriSpeech is not exactly public domain)?

[neonbjb]

Not a dumb question, this is something that took me some pain to figure out. ASR-focused datasets are often poor for TTS because they are missing punctuation and have bad splitting (e.g. not split on sentences). These are both important cues for a TTS system. Both of these applies to LibriSpeech.

I believe LibriSpeech intersects with LibriTTS, so the model should work equally well with voices from either datasets.