πΈ v0.0.15
πBug Fixes
- Fix tb_logger init for rank > 0 processes in distributed training.
πΎ Code updates
- Refactoring and optimization in the speaker encoder module. (π @Edresson )
- Replacing
unidecode
withanyascii
- Japanese text to phoneme conversion. (π @kaiidams)
- Japanese
tts
recipe to train Tacotron2-DDC on Kokoro dataset (π @kaiidams)
πΆββοΈ Operational Updates
- Start using
pylint == 2.8.3
- Reorg
tests
files. - Upload to pypi automatically on release.
- Move
VERSION
file underTTS
folder.
π Model implementations
- New Speaker Encoder implementation based on https://arxiv.org/abs/2009.14153 (π @Edresson )
π New Pre-Trained Model Releases
- Japanese Tacotron model (π @kaiidams)
π‘ All the models below are available by tts
or tts-server
endpoints on CLI as explained here.