Free speech-to-text missed one in four words on medical lecture audio.

Medical English requires exact terminology, but we need to know if automated speech recognition tools transcribe complex lessons reliably. Researchers turned a biochemistry passage into spoken audio using text-to-speech technology. Then, they tested how accurately four online tools could listen to the audio and type it back out. Two free tools captured most of the text, but two paid tools missed large sections of it. Across all the platforms, skipped words were the main source of mistakes, often dropping numbers and figure labels. Although tools rarely swapped in the wrong word, their substitutions replaced crucial medical terms with incorrect scientific words. Because small mistakes distort medical facts, teachers and students cannot fully rely on automatic transcripts without human review.

Why it matters

Medical English students use automatic transcription to review lectures. If the transcript silently deletes numbers or swaps 'imino' for 'amino', a learner memorizes the wrong chemistry. This paper shows that paid services can actually miss more words than free ones, and that coverage—how much of the original speech survives—matters just as much as accuracy for real study.

Questions answered