How Often Does Speech Recognition Get Notes Wrong?
Often enough that every dictated line needs a read. A 2018 study found 7.4 errors per 100 words in raw speech-recognition output, and errors remained in 42.4% of notes after the clinician signed them.
What did the study measure?
Zhou and colleagues pulled 217 notes dictated by 144 physicians at two health systems and counted every error at three stages: raw speech recognition, after a transcriptionist, and the signed note. Raw output carried 7.4 errors per 100 words (JAMA Network Open, 2018). Signed notes still had errors 42.4% of the time.
Where do the errors hide?
Not in the long words. In the short ones. "Fifteen" heard as "fifty." "No" dropped from "no chest pain." Hydroxyzine landing where I said hydralazine. The sentence still reads fine, which is the problem: your eye slides past it at 6 pm.
What should I check before I sign?
Read every number, every negation and every drug name. Those three carry the risk. I built Chartside's dictation to mark a word it didn't catch clearly instead of guessing, so the shaky word is the one you look at, not the one you miss. More on drug names in why dictation mishears them.
If that is how you'd want dictation to behave, the early-access list hears first on December 1.
