Context based multimodal fusion