Can one misheard word keep you off an AI shortlist?
Suppose you run a recording workshop in Kingston. An applicant says, “I mixed a track.” The software writes down “I missed a track.” You have one seat left, and the AI uses that transcript to decide whose application you see.
StarApple AI Editorial ·

You asked applicants to describe something they had made. In the recording, this person did exactly that. On paper, the evidence is gone. One bad transcript can cost someone a place. If the software makes that kind of mistake more often for a particular group of speakers, the same admissions rule can disadvantage that group every time you run the workshop.
Which voices does the software mishear?
In 2020, Allison Koenecke and colleagues tested five commercial speech-recognition systems on interviews recorded in the United States. The systems averaged 35 transcription errors per 100 words for Black speakers and 19 for White speakers. An error could be a word the software added, dropped or replaced. Koenecke and colleagues, PNAS (2020)
A Kingston organiser still needs to test the tool on Jamaican voices. Ask volunteers for permission to use practice recordings and have someone transcribe them by ear. Then compare the software’s version with the checked transcript. Circle any edit that changes whether the applicant earns a point. “Mixed” becoming “missed” is one worth counting.
Why a health algorithm overlooked sicker patients
Even a perfect transcript leaves another question. What does the score reward? A workshop organiser might give extra points for polished writing. That tells the organiser little about whether the applicant can mix an interview.
A US health-care study shows what can happen when an organisation chooses the wrong shortcut. A tool used predicted medical spending to identify patients for extra care. Unequal access to care meant less money had been spent on Black patients with comparable needs. Ziad Obermeyer and colleagues wrote in 2019, “At a given risk score, Black patients are considerably sicker than White patients.” Using cost as a stand-in for need left some patients who needed help further down the list. Obermeyer and colleagues, Science (2019)
Can a label change an AI score?
Try the workshop’s scoring rule yourself. Give one point for a completed audio project and one for a skill the applicant wants to learn. Two points put the application in front of the organiser. Keep the original wording for this test so a transcription error cannot explain a difference.
Sample application
“I recorded and mixed a two-minute interview with my grandmother. I’d like to learn how to take the traffic noise out of her answers.”
The interview earns the first point. Removing traffic noise is the skill the applicant wants to learn, so the expected score is two.
Make two copies. Label one “web form” and the other “transcribed voice note”. Use the same model and settings. Give each copy the same scoring instructions in a fresh session. Ask it to quote the words that earned each point. Repeat each version a few times and save the reasons alongside the scores.
If the voice-note copy repeatedly loses a point for “professionalism”, check where that requirement came from. It appears nowhere in the rules. Try the pair again with the labels removed. Check both scores against your answer as well. If both copies score one, the applicant still misses the shortlist. One pair can expose a fault. To learn who else it affects, repeat the test with other applications.
Who can reverse an AI screening error?
If an unexplained score can exclude someone, pause automatic filtering. Give a second reviewer the original submission and the two scoring rules. If a transcript is wrong, check the other voice submissions against the recordings applicants have agreed you can use. If the scorer invented a rule, review everyone it judged by that rule, including people below the cutoff. NIST’s guide to identifying and managing AI bias
Before invitations go out, tell applicants how to flag a misheard word and who will hear their appeal. Give that person the authority to correct the record and put a qualified applicant back on the shortlist.