that isn't possible currently, the only way you can do that is with a multimodal model so openai's real time api that is still being released