Interviewing
What an Accent Filter Changes, and What It Cannot Fix
No study measures an accent filter against a hiring transcript, so buying one is a bet in both directions. The worry underneath it is real: in a 2019 test of five commercial recognizers, error on identical phrases ran about twice as high for Black speakers. But a filter is one more program between you and the platform, and the recording is still there for a person to play. The measure with evidence behind it is the microphone: a quiet room and a slower pace on names and numbers cost nothing.
The takeA filter asks you to pay for a problem you did not cause. The measured error sits in the recognizer, not in your mouth, and the study that separated the two found the gap holding on identical words. That does not make the tools contemptible or the people who buy them naive. It makes the purchase a private fix for a defect in a tool somebody else chose, which is worth naming before spending anything on it.
Where Olive fits
Open a role and see what the work shows
An Olive assessment is not an interview, and nothing drawn from how you sound enters its findings or reaches the employer as a rating: you work a real assignment openly with an AI assistant, think aloud or type, and a human reviewer writes or edits the findings on what happened, in words. You are granted the same report the employer sees, free.
Assess your shortlistWhat Does an Accent Filter Actually Change?
It changes the audio a platform records, and it stops there. On HireVue, a platform that publishes its pipeline, a transcript is produced from that audio, the scoring runs on the text, and the recording stays available to the people running the round. So a filter touches one step of several. Whether the words coming out of it land more accurately in a transcript is the part nobody has published a measurement of.
A filter is an unmeasured bet, in both directions.
Real-time conversion is also live software running while you speak. It can drop out, add latency, or leave a recording sounding processed, and a platform expecting a plain microphone feed has no reason to handle it gracefully.
How often any of that happens is unmeasured too.
It is the ordinary risk of putting one more program between a microphone and a website.
Accent softening and a voice changer are also not the same purchase. Changing how your own speech is pronounced is a cosmetic layer on your voice. Having a different person give your answers is a claim about who is in the interview, and that is a question about honesty rather than about audio. A synthetic voice reading your own answers sits between the two: tell the employer before the round rather than let it pass as your own voice.
Check the platform's own instructions before adding anything to the audio chain.
Is the Worry Behind the Question Real?
Yes, and it has been measured more than once. In a 2020 audit of five commercial speech recognizers as they stood in 2019, matched short phrases spoken by both Black and white speakers came back with word error rates about twice as large for the Black speakers 1. The authors put that down to the acoustic model, which means pronunciation rather than word choice. The speakers were American, so it measures dialect rather than a second-language accent.
So a more formal vocabulary is not the lever.
A second study points the same way for people who learned English later. Researchers ran recordings of 2,713 speakers reading one paragraph through Amazon, Google and Microsoft in 2021 2. Information lost, on a scale from 0 to 1, was 0.14 lower for speakers whose first language was English, and it rose the later a speaker had started English 2. A single read paragraph is easier than a spontaneous interview answer, and all three services have been updated since.
One platform vendor reports a narrower spread for its own supplier. HireVue names Rev.ai as the supplier producing the transcripts its scoring reads, and publishes internal testing putting word error at 7 to 10% across White, Black, Hispanic and Asian applicants, against 15 to 30% for the alternatives it tested 3.
That is a vendor's own unpublished analysis, with no sample size and no independent replication, so it sits beside the published research rather than settling it.
The hiring side of this site reads the first of those studies alongside others. Its page on them, how accurate AI interview transcription is and whose speech it gets wrong, tells interviewers to keep the audio and judge what a candidate said, not how cleanly it came back.
None of those studies tested a filter.
Speakers already do a manual version of one. In a two-week diary study of 30 African American participants, 93% reported changing how they spoke to be understood by voice technology, and most reported frustration when the errors came anyway 4. Thirty paid participants using consumer devices is not a hiring population.
Fix the Microphone Before the Voice
The audio itself is where a documented failure sits. HireVue's 2025 statement names the ways an answer stops being scored by its AI: a microphone problem, an answer too short, one not given in the time allowed, or speech the system cannot understand clearly enough 5. Those responses go to a person instead 5.
That is one platform's published process rather than an industry rule, and another product may simply score a thin answer low. Flagged for human review is not the same as a good outcome either. But it puts a vendor-stated consequence on the audio that arrives, and the microphone is the part of that audio you can test before the round, which no measurement offers for a filter.
So spend the effort there first.
- A wired headset, or a microphone close to your mouth, tested by recording yourself and playing it back
- A room with the door shut and nothing humming: a fan, a television next door, traffic through an open window
- A pace that slows on proper nouns, company names, tools and numbers, which are the words where one misrecognition changes what an answer means
- The spelling of an unusual name typed alongside the answer, if the platform gives you a text field
That list promises nothing about how a round turns out. It is the part of the setup you control, it costs little or nothing, and it fails in ways you can hear when you play the recording back.
What If the Round Goes Badly Anyway?
Ask what the tool read before drawing a conclusion about your voice. In the United States, the EEOC's 2016 enforcement guidance on national origin discrimination sets a two-part test for letting an accent count 6. The job's duties have to genuinely require effective spoken English, and the accent has to materially interfere with communicating in it 6.
That is the agency's position rather than a statute or a court ruling, and it was written about decisions people make. It does not mention speech recognition, and nothing here tests how a rejection driven by a machine's misreading would be analyzed.
It still gives you a specific question to ask: whether the job's duties require effective spoken English at all.
Put three more questions to the recruiter: what produced the transcript, whether a person reviewed the recording, and what the tool scores. What an AI interview scores and what it does not covers how to find that out for your own round. Where speech itself is the obstacle, whether an AI interview tool can score you down for an accent or a disability covers that ground.
An accent filter is not forbidden, and nobody should feel foolish for looking at one. It is the item on the list with no evidence behind it, paid for privately, aimed at a defect in a tool somebody else chose. The recruiter's answer costs nothing and tells you more.
Common questions
Does an accent softener improve the transcript in an AI interview?
No published measurement answers that. The studies that exist measure how commercial speech recognition performs on real accented speech, not on speech a filter has already processed. So there is no evidence it helps the transcript, and none that it hurts it. What is documented sits at the other end of the chain: one vendor names a microphone problem as a reason an answer is not machine-scored at all. The audio quality is the part with a stated consequence attached to it.
Is using a voice changer in an interview dishonest?
It depends what it changes. Softening how your own speech is pronounced is a layer over your own answers. Having another person deliver them is a claim about who is in the interview, and that is a misrepresentation whatever the audio sounds like. A synthetic voice reading your own answers is something to tell the employer about before the round. Where the platform or the employer states rules about what may sit between a candidate and the recording, read them before the round rather than after it. If an employer asks, say plainly what is running.
Will speaking more formally make the transcript more accurate?
Vocabulary is not what the measured gap is about. A 2020 audit of 2019 versions of five commercial recognizers compared identical short phrases spoken by Black and white speakers and found word error rates about twice as large for the Black speakers, which points at the acoustic model rather than at word choice. Slowing down on proper nouns, company names and figures is a different matter. Those are the words where one misrecognition changes what an answer means.
Can an employer reject me because of my accent?
The EEOC's 2016 enforcement guidance on national origin discrimination, which sets out the US agency's position, says an employment decision may rest on an accent only where two things hold: the job's duties genuinely require effective spoken English, and the accent materially interferes with communicating in it. The guidance also gives an example where rejecting an applicant on accent was lawful, so it is not a rule that accent can never matter 6. It is agency guidance rather than a statute, and it was written about decisions people make, not about machine transcription.
What should I ask if a transcript-scored round goes badly?
Ask what produced the transcript, whether a person reviewed the recording, and what the tool actually scores. Those are answerable questions about the employer's process rather than about your voice, and a recruiter who cannot answer any of them has told you something too. Where speech itself is the obstacle rather than the room, the article on accents and disability in AI scoring covers what a request for another assessment format has to say to count as an accommodation request.
References
- 1. Racial disparities in automated speech recognition pmc.ncbi.nlm.nih.gov The matched-phrase finding this article rests on: on identical short phrases, word error was about twice as large for Black speakers across all five systems, which the authors trace to the acoustic model rather than to vocabulary.
- 2. Global Performance Disparities Between English-Language Accents in Automatic Speech Recognition arxiv.org Supports the second measurement in this article: 2,713 speakers reading one paragraph through Amazon, Google and Microsoft in 2021, with Word Information Lost 0.14 lower for first-language English speakers and rising with a later start in English.
- 3. 2025 Explainability Statement hirevue.com The vendor's own internal transcription testing quoted in this article, naming Rev.ai as the supplier and giving a 7-10% word error range across ethnic groups against 15-30% for the alternatives it tested.
- 4. "I don't Think These Devices are Very Culturally Sensitive." - Impact of Automated Speech Recognition Errors on African Americans frontiersin.org Supports the sentence that speakers already modify how they talk for a machine: 93% of 30 participants in a two-week diary study reported doing so, with frustration reported when errors came anyway.
- 5. 2025 Explainability Statement hirevue.com Names a microphone problem, an answer too short, an answer not given in the allotted time, or speech the system cannot understand clearly, as reasons a response is not AI-scored and is flagged for human review instead.
- 6. EEOC Enforcement Guidance on National Origin Discrimination eeoc.gov The two-part test this article states: an employment decision may rest on accent only where effective spoken English is required for the job's duties and the accent materially interferes with communicating in it.
6 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.