Analyzing Learned Representations of a Deep ASR Performance Prediction Model