Speaker-centric multimodal understanding