Has anyone had great success with voice mode following their system prompts? Every time I try the quality seems markedly less than the normal gpt-4o model.