[Help]: In the S2A model, have you tried applying classifier free guidance on the semantic condition?

[xliu99]

Hi, Thank you so much for the great open source work.

I found that in the second stage S2A model, the classifier free guidance is only applied on the acoustic prompt. I am wondering have you tried applying the CFG also on the conditioning semantic tokens (maybe that will improve WER), or on both the prompt and the semantic tokens?