The trained checkpoints include:
- Hubert
base
(from fairseq, see the fairseq repository and original HuBERT paper for more details), - HiFiGAN trained to vocode HuBERT
base
layer 6 features. - ASGAN HuBERT variant checkpoint. This is the best performing model trained in the paper.
- The validation and test splits used for validating and training the model for the SC09 dataset.