You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hello, I see the training code for the PyTorch implementation of TacoTron2 and FastPitch is covered under BSD-3. I also see the underlying LJ Speech Dataset is not copyrighted and in the public domain. However I am unclear about what the model weights for the pre-trained models are licensed under? In particular, the model weights that are downloaded in the download_dataset.sh script.
The text was updated successfully, but these errors were encountered:
Hello, I see the training code for the PyTorch implementation of TacoTron2 and FastPitch is covered under BSD-3. I also see the underlying LJ Speech Dataset is not copyrighted and in the public domain. However I am unclear about what the model weights for the pre-trained models are licensed under? In particular, the model weights that are downloaded in the download_dataset.sh script.
The text was updated successfully, but these errors were encountered: