Beware of Errors in Voice Recognition Messages
The other day I got a text message reminding me of an upcoming meeting with someone named Morris.
The problem was I didn’t know anyone named Morris, at least nobody I was supposed to be meeting. So I called the sender – who was an attorney and friend – for clarification. Turns out he meant to say “meet me tomorrow,” but his speech recognition app mangled it into “meeting with Morris.”
In the process we both received a cautionary lesson on the pitfalls of speech-to-text software.
Errors in Medical Notes
My little anecdote was harmless and humorless. But the risks for professionals can be all too real.
Consider this study in the International Journal of Medical Informatics, in which researchers examined notes dictated by emergency room doctors at a busy hospital that treated approximately 42,000 visitors per year.
The scientists found an average of 1.3 errors per speech-to-text note. Nearly 15 percent were described as critical errors that could affect patient care – or even cause harm. Even the innocuous mistranslations led to confusion in reading the notes.
The mistakes were divided into eight different categories:
- Annunciation errors
- Added words
- Deletions
- Homonyms
- Nonsense errors
- Spelling errors
- Suffix errors
- Dictionary errors
Annunciation errors were the cause of 53.9 percent of the mistranslations, according to this summary. Next were added words (11.7 percent), nonsense errors (10.9 percent) and problems with homonyms (4.7 percent).
3 Risk Management Tips
- If you’re going to use voice recognition, make sure it’s a proven and reliable program.
- Proof-read and double-check before hitting “send.”
- Avoid using VR for sensitive and confidential messages.
On A Lighter Note
Here are some actual VR gaffes submitted by readers of the website Write Works:
- Said: “Anaïs Nin” Came out as: “On the east and then”
- Said: “Woulda-coulda-shoulda” Came out as: “Would occur to shutter”
- Said: “Catch-up” Came out as: “Ketchup”
- Said: “A couple of FYIs” Came out as: “A couple of FY eyes”
- Said: “Thank God” Came out as: “Think God”
- Said: “And then I ran” Came out as: “And then Iran”
- Said: “Let’s dance” Came out as: “Less dance”
- Said: “Castaic Lake” Came out as: “To stay at lake”
- Said: “I do ramble” Came out as: “I do Rambo”
- Said: “Denims and wools” Came out as: “The Nims and wolves”
- Said: “264 Lane Street” Came out as: “26 Four-lane Street”
- Said: “Foreign voices” Came out as: “For invoices”
- Said: “Colorectal cancer risks” Came out as: “Co-director cancel risks”
Do you use speech-to-text or voice recognition software? Got any good stories?
Sources:
- US National Library of Medicine http://www.healthcarebusinesstech.com/speech-recognition-errors/
- Healthcare Business Tech http://www.healthcarebusinesstech.com/speech-recognition-errors/
- Write Works http://www.writeworks.biz/blog/voxrec/
