Does Your Accent Affect Your AI Interview Score? What the Research Actually Found

First: Speech-to-Text Really Is Less Accurate for Some Voices

In 2020, a team led by Allison Koenecke published "Racial Disparities in Automated Speech Recognition" in the Proceedings of the National Academy of Sciences (PNAS) — one of the most-cited papers on this exact question. The researchers tested five commercial speech-recognition systems (from Amazon, Apple, Google, IBM, and Microsoft) against a matched sample of 73 Black speakers and 42 white speakers, drawn from interviews recorded across five U.S. locations, totaling 19.8 hours of audio.

Every single system performed worse on Black speakers' voices. Averaged across all five, the word error rate was 0.35 for Black speakers versus 0.19 for white speakers — roughly double. The two systems with reported figures in the paper: Microsoft's error rate was 0.27 for Black speakers versus 0.15 for white speakers, and Apple's was 0.45 versus 0.23. The authors traced the cause to the training data itself — these systems learn from large audio datasets that underrepresent African American Vernacular English and other dialectal variation, so they're measurably better at recognizing the speech patterns they saw more of in training.

This paper is specifically about everyday automatic speech recognition — voice assistants, dictation, captioning — not interview-scoring software. But it establishes something real and foundational: if an AI interview tool relies on the same category of speech-to-text technology (and most do, since building a bespoke ASR system from scratch is not something most HR-tech vendors do), the transcript it works from is not equally accurate for every voice.

“Averaged across five commercial systems, the word error rate was nearly double for Black speakers versus white speakers.”

Second: Does That Transcription Gap Actually Change Your Interview Score?

The Koenecke study didn't test interview scoring at all, which left an open and more practically important question: even if the transcript has more errors, does that bias make it into the number a candidate is actually judged on? A 2024/2025 study published in the Journal of Applied Psychology — "Automated Speech Recognition Bias in Personnel Selection: The Case of Automatically Scored Job Interviews," by Hickman, Langer, Saef, and Tay — went and tested exactly that, inside a real automatically-scored video-interview pipeline.

The researchers used two samples of mock video interviews (1,014 and 414 interviewees) and ran each recording through one human transcriber plus several commercial automatic speech recognition services (the ASR tools compared included Whisper among others). They checked for transcription bias against ESL speakers, speakers of color (with Black interviewees analyzed separately), and female speakers — and then, critically, checked whether any of that transcription bias carried through into the machine-learning model's final interview scores.

• The transcription bias was real and showed up in nearly every comparison — the commercial ASR services made measurably more errors on ESL, Black, and female speakers' audio, with the size of the gap varying by which service was used.
• But when the researchers looked at the actual ML-generated interview scores, that transcription bias did not produce meaningful bias in the scores themselves — neither in how well the scores correlated with other measures of performance, nor in the average scores different groups received.

“Transcription was measurably biased against the same groups. The final interview score, measured directly, wasn't.”

Why a Noisier Transcript Doesn't Necessarily Mean a Lower Score

Neither paper spells out a confirmed mechanism for why transcription errors didn't propagate into scoring bias — that's a genuinely open research question — but the shape of the result is worth understanding, with the caveat that this is a reasonable inference, not something either paper states outright.

Interview-scoring models generally aren't grading your interview the way a strict dictation test would — penalizing every mistranscribed word directly. They're typically trained to pick up on broader signals across a whole answer: specificity, structure, relevant content, consistency across follow-ups. A handful of garbled words scattered through a two-minute answer may simply get absorbed into that larger pattern without shifting the overall read of the answer's content, in a way that a strict word-for-word transcript-matching system would not be as resilient to.

That's a plausible explanation for what the Hickman et al. study found, not a guarantee about how every AI interview vendor's scoring model works. It's also not evidence that AI interview tools are bias-free in general — a separate body of research on tone, vocal delivery, and camera-gaze scoring shows other dimensions where bias has shown up. This is specifically about accent and dialect feeding into transcription.

What This Research Doesn't Settle

It's worth being precise about the limits here, because overclaiming in either direction would be dishonest.

• The Hickman et al. study tested specific commercial ASR services on mock interview data — it did not test every AI interview vendor on the market, and no publicly known study has directly audited a named vendor's production scoring model for accent-driven bias the way HireVue's facial-analysis scoring was independently audited and then discontinued in 2021.
• ESL status and racial/dialectal categories are related but not identical to "having an accent" in the colloquial sense — a candidate with a strong regional accent that isn't captured by those category labels wasn't necessarily represented in this specific sample.
• "Didn't produce meaningful bias in this study" is evidence the risk is smaller than the transcription-error numbers alone would suggest — it is not a blanket guarantee that no AI interview tool, anywhere, ever scores accent unfairly.

This sits alongside the broader, honest picture of how AI interview scoring actually works — evidence for specific kinds of bias reduction in specific studies, not a blanket claim that every AI tool is fair on every dimension.

What's Actually Worth Doing About This

Given what the research does and doesn't show, here's where to actually put your energy:

• Don't try to mask or neutralize your accent. There's no good evidence that changing how you sound measurably helps your AI-scored content, and it's a distraction from the thing the research suggests actually drives the score: specific, structured, consistent answers.
• Do control what audio quality you can control. Use a decent microphone if you have one, minimize background noise, and don't rush your answers — not because the AI is secretly grading your accent, but because any speech-to-text system (biased or not) does better with clean, clearly-paced audio, and a cleaner transcript just gives the scoring model more to work with either way.
• Focus your practice on content, not delivery. The research consistently points to specificity, structure, and how an answer holds up under follow-up questions as what's actually being scored — not vocal tone or polish.
• If you genuinely suspect a specific platform scored you unfairly, that's a legitimate thing to raise with the employer or platform directly — this research doesn't rule out any particular vendor having a problem, it just shows the overall risk is smaller than the raw transcription-error numbers imply.

If you're preparing for an AI-moderated interview generally, understanding how one-way video interview scoring actually works is worth reading alongside this.

And if you're wondering what else about the format might be working against you, the research on eye contact and AI video interviews covers a different, equally over-worried-about variable.

Key Takeaways

If you remember only a few things, remember these:

• A 2020 PNAS study (Koenecke et al.) found commercial speech-recognition systems had roughly double the word error rate for Black speakers versus white speakers (0.35 vs. 0.19 averaged across five systems, from a sample of 73 Black and 42 white speakers).
• A 2024/2025 Journal of Applied Psychology study (Hickman, Langer, Saef & Tay) tested this directly inside an automatically-scored video interview pipeline (two samples, 1,014 and 414 interviewees) and found the same kind of transcription bias — but it did not carry through into measurable bias in the final ML interview scores.
• Scoring models appear to evaluate broader patterns across an answer (specificity, structure, consistency) rather than grading a strict word-for-word transcript, which is a plausible — not confirmed — reason the transcription gap didn't become a scoring gap.
• This doesn't cover every AI interview vendor, and it isn't a guarantee no tool anywhere has an accent-bias problem — it's evidence the risk is smaller than raw transcription-error numbers suggest, not proof the risk is zero.
• The highest-value use of your prep time stays the same regardless: specific, structured, well-rehearsed answers — not changing how you sound.