Multi-modal speaker diarization of real-world meetings using compressed-domain video features