Generation of synthetic speech with expressive content