Actually I am about 99% sure that audio transformer models can simulate all sorts of emotions or personality. I mean with Eleven Labs, which is 100% realistic, you could just train it on a voice that was constantly expressing some emotion. Then train another voice ID with the same person but different emotion.

Pretty sure they are working on ways to add inflection or something.

But 100% as far as the text goes, GPT4 can assume a personality or include jokes in interactions. Will it necessarily be as "good" as your text chat personality I doubt it but who knows.