How are you prompting to get the more natural conversation style? Or are you using custom voice models for that?