Spaces:
Running
I believe the model works best when processed sentence by sentence
This truly is a revolutionary model, and I'm really impressed with it. I think the model could benefit from auto-chunking longer texts because otherwise the model breaks down and stops being able to accurately pronounce words.
Thanks for the suggestion. We actually tried that approach before, but we found the results weren't great. The resulting audio didn't sound very consistent, so we decided to remove that feature for this version.
So what's thte best way to generate long/infinite sentences/paragraphs if it doesn't auto chunk?
Also what about streaming?
So what's thte best way to generate long/infinite sentences/paragraphs if it doesn't auto chunk?
Also what about streaming?
There's currently no great way to control consistency across multiple generations. You might try increasing the CFG value to see if that helps. And we plan to improve this in a future update.
For streaming, the latest inference code has been merged into the repository. You can now clone the repo to use it.