Spectral Clustering for Speech Separation