When the AI Interview Mishears Your Accent

If you speak English with an accent the model handles poorly, the most likely way an AI-mediated interview hurts you is not that a model scores you down. It is that a human recruiter later reads a transcript full of errors and rates you — and your answers — worse than they would have rated the audio.

That distinction matters because it changes what you should do about it. The published evidence on automatically scored interviews suggests the scoring models themselves absorbed transcription errors without producing systematically biased predictions. The newer evidence on human readers points the other way. This guide separates the two, says what each study actually measured, and sets out the practical asks that survive both findings. It is current as of October 7, 2026.

Where the transcript comes from

Almost every remote interview format in 2026 produces a text artifact, and each produces it differently:

• An AI interviewer conducts the conversation itself and transcribes your side in real time. Voice agents from Ribbon, Apriora, HackerRank, CodeSignal, Eightfold, and others all work this way.
• An asynchronous video interview records your answers and transcribes them afterward for review.
• A human interview with a notetaker attached records the call and produces a transcript nobody explicitly asked you to approve. Whether an AI notetaker is recording your interview is often not announced clearly.
• A live technical interview may be transcribed by the platform even when a person is asking the questions.

The format determines who reads the text. HireVue's Interview Insights product page describes "searchable transcripts" and states that "AI-generated transcripts allow interviewers to tag relevant experience," alongside "instant summaries" (HireVue, Interview Insights). That is the exposure: the transcript is not an internal intermediate step that gets discarded. It is the document reviewers search, tag, quote, and share. Our breakdown of what HireVue records and how it scores covers the rest of that pipeline.

Does speech recognition really make more errors on some speakers?

Yes, and the size of the gap has been measured in peer-reviewed work.

Hickman, Langer, Saef, and Tay studied automated speech recognition applied to job interviews and published the results in the Journal of Applied Psychology in 2025. They found that transcription errors were most frequent for interviewees who spoke English as a second language, for non-White interviewees, and for male interviewees (Hickman, L., Langer, M., Saef, R. M., & Tay, L., 2025, Journal of Applied Psychology, 110(6), 846–858, doi:10.1037/apl0001247). A summary of the study by I/O at Work reports that OpenAI's Whisper produced the most accurate transcripts of the systems compared, ahead of Amazon Transcribe and IBM's offering (I/O at Work).

A 2026 study in npj Digital Medicine quantified the accent gap in a different domain. Testing Whisper and WhisperX on clinical speech, Fatapour and colleagues measured a word error rate penalty for non-native English speakers of 11.0 percentage points with Whisper (p = 0.031) and 3.4 percentage points with WhisperX (p = 0.07). Post-processing WhisperX output through GPT-4o cut the non-native speaker penalty to 1.7 percentage points (p = 0.057) (Fatapour et al., npj Digital Medicine 9, 297, published March 2, 2026). That study measured medical dictation, not interviews, so treat the magnitudes as an indication of scale rather than a figure for your screening call. The direction is consistent with the interview research.

The part most coverage gets wrong

Hickman and colleagues also tested whether those error disparities propagated into the automated scores built on top of the transcripts. They did not find that the errors systematically produced biased predictions of applicant skills and abilities. That is a real finding and worth stating plainly: on the evidence available, the machine-scored component was more robust to transcription noise than the error rates alone would suggest.

If the story ended there, the practical advice would be "do nothing." It does not end there.

The risk is the human who reads the transcript

In March 2026, Kadoma, Shrivastava, and Naaman published a preregistered experiment with 207 U.S.-based crowdworkers. Participants watched speakers with a range of accents deliver a talk, with subtitles that were either accurate or error-prone. Error-prone subtitles "consistently reduce both speaker and content evaluations for all speakers." The authors did not find disparate impact between accent groups once subtitle quality was held constant, and drew the conclusion that follows: speakers whose accents ASR handles poorly "are likely to be further penalized by viewers with lower evaluations" (Kadoma, K., Shrivastava, P., & Naaman, M., arXiv:2603.15807, March 16, 2026).

Read the two literatures together and the mechanism is specific. The model generating a score may handle a noisy transcript adequately. The person who opens the transcript, searches it, and forms an impression does not. Errors in the text lower their estimate of both the speaker and the substance of what the speaker said — and the people most likely to be handed an error-filled transcript are the people ASR transcribes worst.

Two limits on how far to carry this. The Kadoma study used crowdworkers rating a talk, not recruiters rating candidates, and 207 participants is a short-paper sample. And neither study establishes what happens in a specific hiring process at a specific employer. The inference is well supported as a direction of risk; it is not a measured effect on hiring outcomes.

The legal picture is unsettled, and it is about accommodation

The clearest live dispute in this area concerns accommodation rather than accent generally. On March 19, 2025, the ACLU of Colorado filed complaints with the Colorado Civil Rights Division and the U.S. Equal Employment Opportunity Commission against Intuit and HireVue on behalf of a deaf and Indigenous Intuit employee. The complaint alleges she requested human-generated captioning so she could access interview instructions and questions during a promotion evaluation, was not given it, and was rejected with reference to her "communication style"; it alleges the platform's automated speech recognition performs worse for non-White and deaf or hard of hearing speakers. It claims violations of the Americans with Disabilities Act, Title VII, and the Colorado Anti-Discrimination Act. HireVue's chief executive called the complaint "entirely without merit" and said Intuit "did not use a HireVue AI-based assessment" (HR Dive, March 2025).

Those are allegations and a vendor denial, not findings. What the filing usefully demonstrates is which legal theory has traction: a documented, refused request for an accommodation that would have removed the dependency on ASR. An accent alone is not a protected characteristic in the way a disability or national origin is, and a general complaint that software transcribed you imperfectly has no obvious hook. A written request on the record is a different matter.

Note also where the request has to go. HireVue's candidate FAQ tells candidates to "request accommodations from your recruiter or the hiring team at that company" rather than from HireVue (HireVue candidate FAQ). The vendor runs the software; the employer owns the obligation. Our guide to asking for an assessment accommodation before you start covers the mechanics of making that request land.

What actually reduces transcription error

Most advice in this area is folklore. These are the interventions that follow from how ASR systems work, with an honest note on what is unverified.

Fix the audio chain first. Word error rate rises with background noise, reverberation, and lossy compression before it rises with accent. A wired headset microphone positioned close to your mouth, a room with soft surfaces, and a wired network connection remove more error than any change to how you speak. This is the highest-value intervention and the one candidates skip.

Avoid overlapping speech. Diarization errors — the model attributing your words to the interviewer or merging turns — produce transcripts that read as incoherent rather than merely misspelled. Let the other speaker finish.

Say proper nouns and acronyms once, slowly, in context. Company names, framework names, and initialisms are where ASR fails hardest, and they are exactly the keywords a reviewer searches the transcript for. "Kubernetes — the container orchestrator" gives a human reader a recoverable path even if the first word is mangled.

Do not try to flatten your accent. There is no evidence that consciously altering your accent mid-interview improves ASR accuracy, and effortful speech modification costs attention you need for the actual answer. Speak at a measured pace and finish your words; that is the part that is supported.

Ask for the transcript. Several jurisdictions give you a data access right that reaches the transcript of your own interview. Reading it is the only way to learn whether this problem is affecting you at all.

What to ask for, in writing

If you have reason to think ASR handles your speech badly — a deafness or speech difference, or direct experience of garbled captions — make the request before the interview, in email, to the recruiter:

1. A live human interview instead of an automated one, or a human reviewer who watches the recording rather than reading the transcript.
2. Human-generated captioning of the questions, if accessing them is the barrier.
3. Extra time, which reduces the rushed, overlapping speech that degrades transcripts.
4. Confirmation of who reviews what — specifically whether a human watches the video or only reads the generated text.
5. A copy of the transcript, and a route to flag errors in it.

Keep the thread. A dated email asking for something specific, and the reply, is the record that matters later.

If you have already taken it

Request the recording and transcript under whatever access right applies to you, and compare them. If the transcript materially misrepresents what you said, write to the recruiter identifying specific errors with timestamps and ask that the recording be reviewed instead. That is a narrow, checkable request. A general objection to AI interviewing is not, and it tends to end the conversation.

Where this leaves you

The defensible summary as of October 2026: speech recognition demonstrably makes more errors on some speakers' speech; the automated scoring layer appears to tolerate those errors better than expected; the human reading the transcript does not. So spend your effort on the audio path, on speaking in a way that keeps searchable keywords recoverable, and on a written request for a review route that does not depend on a machine transcript — in that order.

One thing worth testing on the other side of the same problem: any assistant that transcribes the interviewer for you is running the same class of ASR and inherits the same failure modes, including on the interviewer's accent and on a noisy conference-call audio path. Control is a desktop AI interview assistant for Windows and macOS whose live transcription works on the interviewer's speech, and its free allowance — five messages and two minutes of voice, no card required — exists to be spent finding out whether it holds up on your actual setup. Our guide to testing an AI interview assistant before an interview describes a test worth running before it matters.