Models that map spoken language to objects in an image would make it easier for customers to communicate with multimodal devices.