Here's what the Batch Inference part does:
- Try to put as much characters as possible into one batch (200 max)
- If it's not possible, it'll try to do a cut whenever there's a semicolon character
- If it's not possible, it'll try to do a cut whenever there's a comma character
- If it's not possible, it'll try to do a cut after the most logical word (thus, therefore etc.) --> There's a list at the top of the Gradio script, and it's possible to modify it in Advanced Settings
- If nothing above worked, it's just going to go past that 200 line (realistically, if your text isn't gibberish, this shouldn't happen :D)
The Podcast Generation feature has these features built in:
- Takes two reference speeches and two reference texts (or empty and then transcribed automatically)
- You have to give a name to each of the two speakers
- You can then paste the podcast script, with one speaker's name followed by a semicolon and then their text, you can do the same with the other speaker, all as long as you want (because it's using the same batch inference as before)
All in all, the batch inference feature allow for a little bit more than real-time inference. (I might do another pull request with real-time streaming)
Immense thanks to all of those who worked on this project, it's really great. There's of course still room for improvement, but I think this is a step forward in terms of OSS TTS, so thanks !
