The use of large language models has made it easier for dictation software to quickly deliver better results. Is that enough to push the technology into common use in the office?