Few-shot learning support?

Amazing work, the quality and diversity of the demos are so good!
I just want to know if this model supports few-shot learning. When inferring, aside from text descriptions, can I add a prompt audio to guide the generation process? 