Control speaking manner
I am looking at the api documentation on the Breeze site. I cannot figure out how to control vocal qualities. For example 'whispering'.
I see the 'instruction' key is json api but its unclear what the model actually supports. I can't find relevant audio examples on the site.
@sirovub
sorry about the confusion.
the instruction supports natural language, in practice, it supports both short and long phrases, and it understands the instruction pretty well.
so there's not a limited set of tags, but anything you want it to say like.
in your case, you can simply type "whispering" in the instruction to see how it works (cfg value=2-4 usually helps).
For what qualities the model can support, you can see our in-house voice direction benchmark:
https://breezeblue.ai/benchmarks/tts-voice-direction
Thanks! I will give it a spin after legal reviews the privacy policy.
@sirovub
please do give us a comment / email if you have any concerns wrt our privacy policy, we are in the process of revising it.
@sirovub
sorry about the confusion.
the instruction supports natural language, in practice, it supports both short and long phrases, and it moviebox understands the instruction pretty well.
so there's not a limited set of tags, but anything you want it to say like.
in your case, you can simply type "whispering" in the instruction to see how it works (cfg value=2-4 usually helps).
For what qualities the model can support, you can see our in-house voice direction benchmark.
I appreciate that you’re open to receiving comments or emails regarding the privacy policy, especially while it’s currently being revised. Privacy is something users should be able to understand clearly, so I’d definitely encourage everyone who has concerns or suggestions to share them directly. Hopefully, the revised version will make the policies and how user information is handled even clearer and easier to follow.
Also, it’s good to know that the instruction supports natural language rather than being restricted to a fixed set of tags. Being able to simply write something like “whispering” and have the model understand the intended direction makes the feature much more practical and flexible. I’ll definitely give it a try with the suggested CFG range and explore the voice direction benchmark to get a better sense of what qualities it can support.