Stochastic speaking styles and unpredictable uh's & umm's

[RahulBhalley]

I admit that Bark is realistic speech sounding.

But there are couple of issues as following. Could someone please help me fix them?

• The output speech creates so many uh's and umm's even if none is present in the text.
• The speaking style and (sometimes even) voice are different for the same speaker (passed to history_prompt argument of semantic_to_waveform() function) and also vary across different inferences. This reminds of Stable Diffusion which has generator argument to make outputs predictable. It'll be great to have such argument.
• Sometimes there are a long pause before the sentence is spoken. (I am running long_form_generation.ipynb which splits the sentences using NLTK to avoid abruptness in long speech).

[mak1168]

i face the same problem, how could minimzae the um...uh.. as low as possible