Good points, thank you! I just tested it. While ChatGPT was very good in adding generic (textual) annotations, the result for generating SSML where very poor (lack of voice names, lack of distinction between narrator and character etc).

Probably the results with a model trained for this plus human audit could lead to very good results.