Identifying Speaker State from Multimodal Cues