New Finetuning of your models
New API endpoint
Full documentation (online and built in)
A couple of minor file updates.
Small bug fix on file generation in instruct mode with no character name available (thanks to @danielwburch)
Much older systems with low system ram are taking 70 seconds to start up (so Im told). Ive added an extra 20 seconds onto the 60 seconds for subprocess start-up.
documentation updated too.
Had to update a variable due to the latest version of pydantic using a similar variable and causing a potential clash. As such, I have had to use a new configuration file (simplest method). Peoples base settings will be reset on next update. Added a option in the config file for standalone mode output.
Added:
- new documentation
- demo in the documentation
- extra API endpoints (yet to be documented)
- voice notification on change of model or setting
- DeepSpeed update to v11.x and re-documented
- improved vocab bringing the XTTSv2 model closer to sounding like the sample, with improved pronunciation and intonation.
On new chat, character paragraphs, found an outlier possibility where character/narrator wouldn't swap, so performed replace('"<. *', '"< *"').replace('< *"', '< *') which should capture it.
The narrator and filtering has been re-written to vastly improve filtering of text, resulting in less unwanted sounds and better control over narrator/character.
Discards any text to be generated that is 0 or 1 characters in length after the sentence is split down (usually this is a * or a " and just produces a strange sound).
Many of the LLM's aren't using asterisks to correctly split the narration from the spoken text. Have re-written the entire text pre-processing engine to take text within asterisks as being narrated text, text within quotes as the characters voice and to assume that anything else is narrated text (the bits that the LLM's aren't correctly using asterisks with). Also added some additional filtering to clear up some strange text characters that may be sent to the TTS engine.
erew123