Evaluating Speech Synthesis by Training Recognizers on Synthetic Speech