Fast Speaker Diarization Using a Specialization Framework for Gaussian Mixture Model Training