It feels like the voice clone is not using the same latents across threads. Just from how the voices sound, this is most noticeable when there is supposed to be an accent and then one of the generations just drops it completely.
I didn't notice this behaviour so far. Could you please explain what you mean with "across threads"? Can you give a code example where this happens or try to describe your workflow as exactly as possible?
It feels like the voice clone is not using the same latents across threads. Just from how the voices sound, this is most noticeable when there is supposed to be an accent and then one of the generations just drops it completely.