Towards an Optimal Representation of Tones