Robust unsupervised speaker turn detection