Towards Structured Approaches to Arbitrary Data Selection and Performance Prediction for Speaker Recognition