Not All Errors Are Created Equal: Evaluating The Impact of Model and Speaker Factors on ASR Outcomes in Clinical Populations.