What's Changed
- XTTS v1.1 by @WeberJulian in #3089
Full Changelog: v0.17.10...v0.18.0
XTTS v1.1
This model is trained on top of XTTS v1, using output masking. We mask the part of the output that is used as the audio prompt while training and don't compute loss for that segment. This helps us to resolve the hallucination issue that V1 experienced.
Changes
- Add Japanese
- Resolve the hallucination issue (repeating the audio prompt)
- Increased expressivity
- Hash check to control model version
- Added
ne_hifiganthat was trained without denoising that brought some EQ and compression profile that might be unwanted for some use-cases