Towards Representative Subset Selection for Self-Supervised Speech Recognition