By automated translation quality checks followed by manual inspection, and by evaluating baseline ASR, MT, and speech translation models.