Does an AI Interview Understand Accents? Non-Native English Guide

HireVue and Braintrust say their AI scores what you say, not how you sound: your speech becomes text and the text is scored. On that description the risk is a wrong word in the transcript. Every published test we found shows speech recognition making more errors on non-native English, and on varieties such as Indian or Nigerian English, than on US English. In the most recent test the best systems still got about one word in twenty wrong on short read sentences. Take the platform’s practice interview on the microphone you will really use, and listen back or read the captions where the platform lets you.

Platform facts come from the public documentation of Mercor, HireVue and Braintrust, checked on 2026-10-06. Research findings come from the papers in the sources and describe the systems tested at the time. Offermic is not affiliated with any platform named here.

Your words are scored as text

HireVue’s 2026 explainability statement describes three stages: your speech is transcribed, a language model reads the transcript, and the answer is scored against expert human ratings of similar answers. It says the AI “relies exclusively on what candidates say” and does not analyze facial expressions, body language or tone of voice. Braintrust says its AIR interviewer analyzes “your words, not your appearance, tone, or background.”

Mercor says its AI interviewer generates a transcript of your responses and evaluates your performance. It does not say whether anything besides the transcript is used, and it notes that some interviews test language skills.

Where the text is what gets scored, a misheard word is scored as if you said it: “we cut churn” transcribed as “we cut church” (our illustration, not a recorded case) is a different sentence. For phrases, see our guide to a job interview in English.

What the studies measured

None of these studies tested an interview platform; they tested the kind of speech recognition software platforms rely on. They use different error measures, so compare within a row, not between rows.

Study	Systems tested	Speech used	Finding
DiChristofano and colleagues, preprint, 2022	Amazon, Google and Microsoft speech APIs, run in 2021	2,713 speakers born in 171 countries reading one paragraph (the Speech Accent Archive)	All three did significantly better for first-language English speakers. On a 0 to 1 scale of information lost, the average gap was 0.14.
Sanabria and colleagues, ICASSP 2023	Whisper (large model), one unnamed commercial system, wav2vec 2.0	About 40 hours of video-call conversations in many varieties of English	Whisper’s word error rate was 19.7% on these calls, against 2.7% on clean read US English. All models dropped on Indian, Jamaican and Nigerian English.
Graham and Roll, JASA Express Letters, 2024	Whisper, version 3	The same archive paragraph in 18 accents, plus exam conversations	Overall, native accents were transcribed more accurately, though Swedish and German speakers beat British ones. Vietnamese and Thai speakers (22 and 15 people) had the most errors. Conversation was harder than read speech.
McGuire, preprint, 2025	AssemblyAI Universal-2, Deepgram Nova-2, Rev.ai V2, Speechmatics Ursa-2, Whisper large-v3	24 non-native speakers, four from each of six language backgrounds, reading short sentences of common words	Whisper and AssemblyAI averaged 5.4% and 5.6% errors. Four US control speakers averaged 0.7% across systems; the four Vietnamese speakers 14.3%. Hindi speakers had the lowest error of the non-native groups.

How we read this:

• The best recent result is good, on easy material. One word in twenty is close to the error rate the same paper cites for human transcribers, but McGuire’s sentences were short and used only common words. On a longer paragraph, Graham and Roll report 11.5% for women and 14.0% for men from the same Whisper version, native speakers included, on recordings they note may have been in its training data.
• Indian English is not settled. Sanabria found a drop on Indian English in conversation. McGuire’s Hindi speakers had the lowest error of his non-native groups, though still several times the US rate.
• Groups are small. Graham and Roll found fewer errors for speakers who started English younger, so your result can sit far from your group’s.

Rev.ai, the supplier HireVue names, was one of McGuire’s five systems: worst on read sentences (8.6%), best on longer spontaneous recordings from 22 speakers (6.3% keeping filler words and repetitions, 10.3% dropping them). HireVue does not say which version or settings it runs.

What the platforms say

	Mercor	HireVue	Braintrust AIR
On accents	Nothing published; speech system not named	See below	Says AIR is “designed to understand a wide range of English accents”; no figures
Asking for a repeat	The FAQ says to ask the AI to repeat or clarify	No live interviewer; the question is shown in writing or as a video	Not stated
Practice first	Free, unlimited, not shared with companies	Practice questions you can play back	Not stated
Second attempts	Up to three retakes	Set by the employer: 1 to 5 recording attempts per question, or unlimited; HireVue recommends three in total	“Typically” one chance per question

Only HireVue names its speech system and reports numbers. Its 2026 statement says transcription comes from Rev.ai, and that in HireVue’s own tests English word error rates in the United States averaged under 10%, against 15 to 25% for the other systems it tried. It says Rev.ai also beat the alternatives when native speakers were compared with non-native speakers “with a variety of accents” but gives no figures.

An older statement, dated 2022 and still linked from HireVue’s legal hub, did: 12% word error for a Canadian accent and 22% for participants from China with Rev.ai, against 17% and 45% for the alternatives. Those are 2022 numbers for a system HireVue says keeps being improved.

Four other points from the 2026 statement:

• Unintelligible answers go to a person. If the system cannot detect enough content, including when a candidate fails to “speak clearly enough to be understood” or has “technical issues with their microphone input,” the response is “flagged for human review” instead of scored by AI. The 2022 text says this typically happens when fewer than five words are understood, so it will not catch an answer with a few wrong words.
• You can opt out of AI evaluation. Recruiters then rate the same recordings on the same rating scales, and HireVue says opting out does not exclude you from the hiring process. We found no public data on whether human review treats non-native speakers better or worse, so this is an option, not our recommendation.
• Preparation time is an employer setting. The statement lists 0 to 5 minutes or unlimited and recommends one minute. Use whatever you get to re-read the question.
• Accommodation requests go to the employer. The statement’s example is extra time for a candidate with a disability; it describes nothing for second-language speakers.

Mercor’s FAQ adds that interviews are in English unless otherwise specified, that captions are on by default, and that the AI interview is its “standard and only method for screening candidates.” Our Mercor page and HireVue page cover each format in full.

Four things to do differently

1. Use the practice run as a hearing test

No published number tells you how a platform handles your voice on your microphone, so the practice run is the closest thing to a test. During HireVue’s practice questions you can switch microphones or adjust the volume.

Mercor does not say whether the captions show your words or only the interviewer’s, and it says it cannot share transcripts. If your words appear, note the ones that come out wrong, then say those more carefully or swap them for simpler ones.

Offermic, the app we make, runs spoken mock interviews in English and shows live captions of what it heard you say, so you can spot words a speech system gets wrong. It may not use the same speech system as the platform you will face, so treat it as rehearsal, not prediction.

2. Fix the audio before you worry about your accent

Mercor recommends headphones and an external microphone. The accent studies did not test microphones, so the rest is our view: use a headset in a quiet room, and listen to a recording of yourself on it first.

3. Do not rush, and finish your sentences

The evidence supports not rushing; it does not support crawling.

Goldwater, Jurafsky and Manning (2010) analyzed the errors of two recognizers on US telephone conversations; accent was not part of the study. Errors were lowest at typical speaking rates and rose at the extremes, sharply for fast speech. How long each word lasted mattered more than overall rate: words given more time than average were recognized best. The first word of an immediate repetition, and words just before a broken-off word, had error rates 10 to 14 points above average.

Among human listeners, Munro and Derwing (2001) found the best-rated second-language speech was somewhat faster than such speakers usually talk; very fast and very slow speech were rated lower.

Neither study tested a current system. Our view: a little slower than your fastest natural pace is fine; half speed is not. Give each word its full length, and finish a sentence that goes wrong instead of breaking off mid-word. On Mercor, fill thinking time aloud (“Let me think about the best example”): its FAQ warns that extended silences may make the AI interject.

4. Say rare words once, then in plain language

The same 2010 study found that the less probable a word was, the more often the recognizer got it wrong. Those are often the words that carry your answer: product names, tools, acronyms, your employer’s name. Our view: say the name once, clearly, then say what it is in ordinary words, so the meaning survives if the name does not.

This example was written for this post; it is not a real candidate or project.

What nobody publishes

None of the three platforms publishes transcription accuracy by accent for the system it runs today. Nobody says whether the scoring step treats second-language grammar differently: a transcript can be word-perfect and still read as non-native, and the bias checks HireVue describes are by gender, age and race or ethnicity, not first language.

For what AI interviewers assess beyond the transcript, see how to pass an AI interview.