Multimodal analysis of free-standing conversational groups