Voice Interviews and Accents: How Big Is Speech Recognition Bias, and What Safeguards to Ask For

Voice interviews solve a real frontline problem. An applicant for a warehouse, cleaning or security role can answer on their own phone, in the evening, without writing a cover letter. But the answer reaches the recruiter through software that turns speech into text, and that software does not hear everyone equally well. HR directors, works councils and compliance leads are right to ask what happens to the candidate who is misheard.

This article sets out what the research shows, the safeguards that reduce the risk and five questions to put to any voice screening provider, including us. Our earlier piece, No Hiring Tool Removes Bias. Including Ours., raised the accent problem in a single paragraph. How we approach oversight, transparency and record keeping is set out on our trust page.

How big is speech recognition bias for different speakers?

Accuracy varies by speaker, and the published gaps are large.

Researchers measure accuracy with the word error rate: the words a system misses, mishears or adds, as a share of the words actually spoken, compared with a careful human transcript. A rate of 0.35 means about 35 errors for every 100 words.

Four studies set out the problem.

Two points matter for a buyer.

First, much of the cause sits in the systems, not in the speakers. Koenecke and colleagues found the gap was just as large when Black and white speakers said identical phrases, which points to the acoustic models rather than to word choice. Feng and colleagues found that pronunciation differences explained only a fraction of the bias they measured. Asking candidates to speak more clearly does not fix a model.

Second, none of these studies tested a hiring product, and commercial systems are updated over time. That cuts both ways. You cannot assume a provider has the same gaps, and you cannot assume it has none. In our view, the only way to know is a measurement in the languages and for the speakers you actually hire.

Why does a transcription error matter in screening?

Because summaries, evidence and decisions are usually built on the transcript.

In most voice screening, the recording becomes text, the text becomes a summary, and the summary becomes evidence against the role's requirements. A busy recruiter reads the summary first, so an early error travels to the decision unless someone stops it.

Illustrative example. An applicant for an order picker role says: "I have a reach truck certificate, valid until 2028." The transcript reads "rich truck certificate, valid until 2018". The summary records the certificate as expired, and the requirement shows as not met. Nothing in the text looks wrong. Only someone who plays the answer hears what was actually said.

The errors that matter most in frontline screening are small ones:

The risk is highest in three situations. The first is an automated rule that rejects or sets aside candidates based on transcribed text. The GDPR gives candidates the right not to be subject to a decision based solely on automated processing that significantly affects them, such as an automatic rejection, unless an exception applies (GDPR, Article 22). The second is a summary that hides uncertainty: a clean paragraph reads as fact, even over poor audio. The third is overreliance. The EU AI Act requires high risk hiring tools to be built so that the people overseeing them can watch for overreliance on the output and override it (AI Act, Article 14(4)); for recruitment systems, those duties apply from 2 December 2027 (Regulation (EU) 2026/1744).

Accent carries a further risk, because it can signal where someone comes from. Dutch equal treatment law prohibits direct and indirect distinction on grounds that include race and nationality when a job is offered, when a vacancy is filled and in job placement (Algemene wet gelijke behandeling, Articles 1 and 5(1)). Whether a given process crosses that line depends on the facts. In our view, it is a strong reason to make sure no system assesses accent, and to check that transcription errors are not quietly working as an accent filter.

Which three safeguards reduce the risk?

A person hears the original, candidates can switch route, and you monitor.

None of these removes speech recognition bias. Together they stop a transcription error from becoming a hiring decision, and a buyer can check each one in a demo.

1. A person reviews the original answer, not only the transcript

The recording is the evidence; the transcript is a way to find your place in it. Reviewers should be able to play the answer behind any summary line, ideally at the point linked to each requirement, and they should do so before marking a requirement as not met or an answer as unclear. A named person makes the decision after seeing the evidence. An unclear or missing answer goes to that person for a further question, never to an automatic rejection: missing is not the same as failed.

2. Candidates can choose another route

Tell candidates at the start that the interviewer is automated, that a person reviews the result and how to reach a person instead, for example a call with a recruiter or written answers to the same questions. Choosing that route must not count against them. Since 2 August 2026, the AI Act has also required providers of systems that talk directly with people to make sure those people are told they are dealing with AI, unless that is obvious (AI Act, Article 50(1)).

The need is not hypothetical. In a large field experiment with about 67,000 eligible applications for customer service jobs in the Philippines, 7% of interviews led by the AI were aborted because of a technical failure, and in 5% of interviews the applicant said they did not want to continue with the AI (Jabarian and Henkel, 2026, working paper, not peer reviewed). A process built for volume needs a lane for those people.

3. You monitor where transcription fails

Track, per interview language and per site, how often reviewers correct a transcript or overrule a summary, how often interviews break off, how often candidates switch route, and how progression rates differ between interview languages. Each month, take a small sample of interviews in each language and compare the transcript with the recording. Feed recurring errors back to the provider: certificate names, local place names and shift terms tend to repeat. Measuring outcomes by origin raises its own privacy questions, so start with what you already hold, such as the interview language, and take advice before collecting more.

What adjustments should candidates be able to ask for?

Practical changes that let candidates show what they can do.

In the Netherlands, the Equal Treatment (Disability or Chronic Illness) Act, known as the WGBH/CZ, says the ban on discrimination includes a duty to make effective adjustments where they are needed, unless they would be a disproportionate burden (WGBH/CZ, Article 2). The ban applies to offering a job and to handling a vacancy (Article 4(a)). How that applies to a particular voice process depends on the facts, so take legal advice on your own setup.

For a voice interview, useful adjustments include:

Say at the start how to ask, and do not require a medical explanation. Name who handles requests and how quickly. Keep the questions and criteria the same on every route, so answers stay comparable. Record that an adjustment was made, not why: health details are sensitive and rarely needed for the hiring decision.

What should you ask a voice screening provider?

Five questions, in writing, with a good answer for each.

Question 3 has a legal edge as well as a fairness one. Since 2 February 2025, the AI Act has banned AI systems that infer people's emotions from biometric data, such as facial expressions or voice, in the workplace, except for medical or safety reasons. The European Commission's guidelines, which are not binding, say this ban also covers candidates during recruitment (Commission guidelines on prohibited AI practices, C(2025) 5052 final).

For the wider set, covering validity evidence, data and the AI Act, see How to Choose AI Recruitment Software: 10 Questions to Ask Any Provider.

How do you brief reviewers on transcription limits?

Tell them the transcript is a draft and the recording is the evidence.

Since 2 February 2025, the AI Act has required organisations that use AI systems to take measures on AI literacy for the staff who operate them. Since the Digital Omnibus took effect on 27 July 2026, the duty is to take measures that support staff AI literacy, with no set level required for each individual (AI Act, Article 4, as amended by Regulation (EU) 2026/1744). A one page briefing on transcription is a practical part of that:

Our piece on human oversight in hiring covers what makes a reviewer's check real.

What does this mean for staffing agencies and frontline employers?

The safeguards have to hold across languages, sites and clients.

Staffing agencies interview in several languages for several clients at once, and recruiters reviewing for many roles have the least time to listen. Linking each requirement to the moment in the recording that answered it saves that time. Agencies can also use the five provider questions when a client asks how a candidate was screened, and make the alternative route part of standard intake.

Frontline employers face a different trap. Where a role genuinely needs Dutch on the work floor, write that level down as a requirement and test it the same way for every candidate. Otherwise transcription quality can become an informal language test that nobody chose. Works councils will ask what is recorded, who listens, how long recordings are kept and how candidates can switch route; have the answers in writing before the conversation. Our guide to the EU AI Act and recruitment sets out the wider timetable.

Conclusion: the transcript should never be the last word

Speech recognition bias is well documented for some accents, for non native speakers and for people with speech conditions, and no provider can promise it away. What a buyer controls is what happens next: a person who hears the original answer, a route out of the automated interview with adjustments on request, and monitoring that shows where the system mishears. Ask for measured error rates in your languages, prefer an honest "not yet" to a vague claim, and check the safeguards in the demo.

How Radius helps

Radius Hire is an AI hiring platform for frontline roles in the Netherlands. Its AI Interviewer runs a short, structured voice interview that candidates can take on their phone, whenever it suits them, in the language set for the role.

How this fits the AI Act is set out on our trust page.