Each speech-to-text model transcribed the same clinical consultation audio. Outputs are scored by word error rate against the human reference transcript, with medical-term accuracy tracked alongside. The recommended model is the cheapest one that clears the accuracy bar.