A new benchmark finds commercial text-to-speech models mispronounce recently approved medicines far more often than familiar ones.