Can AI interviews understand accented English in customer support hiring?

Updated 4 November 2026

If you lead talent acquisition for a BPO or customer-support operation, an AI interview has to do more than produce a transcript. It needs to understand applicants across the accents and languages in your hiring pool, assess how they handle a customer conversation, and let them complete the process on the devices and connections they have. A misheard product name, order number or escalation detail can leave recruiters reviewing a transcript that does not reflect what an applicant said; if that error feeds into scoring, it can distort the shortlist. Get the assessment or candidate experience wrong and you can also lose good applicants or miss the communication skills customers need. Here is what the research can, and cannot, tell you, and how to test a platform before using it at scale.

Keep three checks separate. First, did the system hear the candidate accurately? Second, does the interview assess skills that matter in the support role? Third, can applicants complete it without avoidable friction? A strong result on one does not prove the other two. Test recognition with representative applicant audio, assessment with a role-specific scorecard, and completion with a live pilot.

First, check whether speech recognition works on your applicants' voices

In a post dated 30 September 2021, HireVue says its transcription word error rates are "in the single digits", meaning fewer than 10 spoken words in 100 wrong. It credits a third-party service, Rev.ai. It says a 20 to 25% error rate was the industry standard "just a few short years ago". It also says it tested native and non-native English speakers with various accents and found the system "extremely robust".

That claim is not enough to predict performance in your hiring pool. HireVue does not publish the test audio, speaker count, date of the result or error rate for individual accent groups in that post. An overall figure can hide differences between groups, and it cannot tell you how the system will handle the voices, names, locations and customer-service terminology your recruiters hear every day.

What the often-quoted 12–22-point gap does, and does not, tell you

You may see a 12–22 percentage-point gap described as proof that speech recognition penalises non-native English speakers. The range comes from a real study, but it is often applied to the wrong population. It is not a study of Indian or Filipino accents, or of first-language differences.

Koenecke and colleagues, first published in PNAS on 23 March 2020, tested 2019-era speech services from Amazon, Apple, Google, IBM and Microsoft on 73 Black and 42 white speakers in the US: 2,141 audio snippets per group, 19.8 hours in all.

Averaged over the five systems, word error rate was 0.35 for Black speakers and 0.19 for white speakers. Microsoft did best, at 0.27 against 0.15, a gap of 12 percentage points. Apple did worst, at 0.45 against 0.23, a gap of 22 points. Those are the two ends of a "12 to 22" range. But the study is about speakers of African American English, it tested cloud systems of that time, and it does not look at first language at all.

What research on non-native English can tell you about recognition errors

The most direct recent measurement is a March 2025 preprint by Michael McGuire, version 2 of 30 March 2025. It tests five ASR systems (four proprietary services and Whisper, an open-source model) on L2-ARCTIC, a research corpus of 24 speakers with six first languages, two men and two women each, reading 2,400 sentences aloud in a lab.

The practical lesson is not to choose a platform from a published leaderboard. Research tells you which questions to ask; only a test using speech like your applicants' can show whether the system handles your hiring context. For a closer look at how Fabric scores spoken English, read Fabric's spoken-English scoring guide.

A gap exists, and it varies a lot. US-English speakers had lower error rates than every non-native group, and Hindi-first speakers had lower rates than every other non-native group. The differences were statistically significant (p < .001). The largest estimated difference in mean match error rate was between Vietnamese-first and Hindi-first speakers, at 0.114.

Two limits apply. These are laboratory read sentences, not interviews. And the corpus has no Filipino-first speakers, so it says nothing direct about Philippine English.

The paper also tests speech closer to an interview: 22 speakers telling stories for 27.5 to 235 seconds each, 26 minutes in all. Mean match error rate ran from 0.063 for RevAI with disfluencies such as “um” retained, to 0.142 for Whisper with disfluencies omitted. Every engine recorded a lower match error rate when disfluencies were retained in the transcription. First language had no significant effect in that set, and the set has no US-English group, so no non-native penalty can be worked out from it.

An Indian-English benchmark makes the local gap more concrete. The 2023 Svarah study contains 9.6 hours of transcribed English from 117 speakers across 65 locations in India. It includes read and spontaneous speech and tests six open-source models plus two commercial systems. The commercial systems recorded word error rates between 20.7% and 30.0%. This is not a result for Fabric's interview engine or your applicant pool, but it shows why a generic English-accuracy claim is not enough for an Indian support funnel.

When can a speech-recognition error affect a hiring outcome?

A speech engine mishearing a word is not automatically a hiring penalty. The error matters if it changes the transcript, influences an assessment and affects the decision. Your team needs to know whether the score depends on a transcript, whether the original recording is available and whether a recruiter can review an answer that appears inconsistent with the candidate's speech.

Do not treat a vendor accuracy claim as proof that it fits your hiring pool

Before relying on an accuracy claim, ask what was recorded, how many speakers took part, which language groups were represented, when the test ran and which metric was used. Then compare it with your own representative audio. A relative improvement is not an error rate, and a result on read sentences may not predict performance in an unscripted customer conversation. Ask the vendor to explain what happens when a misheard word could affect a candidate's score.

Accessibility and accommodation are part of a fair hiring process

On 19 March 2025 the American Civil Liberties Union, Public Justice, Eisenberg & Baum LLP and the ACLU of Colorado filed charges with the Colorado Civil Rights Division and the US Equal Employment Opportunity Commission against Intuit and HireVue. They were filed on behalf of a woman identified as D.K., who is Indigenous and Deaf, and allege violations of the Americans with Disabilities Act, Title VII of the Civil Rights Act of 1964 and the Colorado Anti-Discrimination Act.

According to the charges, D.K. worked seasonally at Intuit across five tax seasons, first as a Tax Associate and later as a Tax Expert Lead, before applying for a Seasonal Manager role. The complaint says she requested human-generated CART captioning for the HireVue interview. Intuit told her the software included subtitles, but she found no subtitle option and used Chrome's automated captions instead. Separately, the complaint describes an earlier Intuit performance metric that allegedly misrecognized her speech and marked her as deviating from scripts. It does not publish a score from the HireVue interview itself.

Both companies deny it. As HR Dive reported, HireVue's chief executive, Jeremy Friedman, called the complaint “entirely without merit” and said Intuit did not use a HireVue AI-based assessment in this instance. Intuit also called the allegations “entirely without merit” and said it provides reasonable accommodations to all candidates. As of 3 November 2026, no public agency determination or ruling has been reported.

For a support hiring team, the lesson is practical: a speech-derived score should not become an unreviewable decision, and candidates need a clear route to request an accommodation or human alternative. This complaint is not evidence of error rates for Indian or Filipino English, and the allegations have not been established by a ruling. It does show why recruiters should be able to review the evidence and handle accessibility needs before a process is rolled out at scale.

Will applicants finish the interview you send them?

Greenhouse's 2026 survey is a signal that candidates' experience with AI interviews deserves attention. It does not give a completion rate for Indian or Filipino support applicants, or a breakdown for customer-service roles, so do not use it as a local benchmark. Use it as a reason to measure your own funnel.

Track invited, started, completed and recruiter-reviewed interviews separately. Tell applicants before they begin that AI is involved, what the interview assesses, whether it records them, who reviews the result and how to request an accommodation or human alternative. If completion drops, investigate the invitation, device, connection and interview experience before assuming candidates lack interest.

What a large Philippine customer-service hiring trial tells you, and what it doesn't

The Philippine evidence is a large randomised test in entry-level customer-service hiring, a useful comparison for teams screening support applicants at volume. Brian Jabarian and Luca Henkel worked with PSG Global Solutions, a subsidiary of Teleperformance. Their working paper, version 2 of 11 September 2026 and pre-registered as AEA RCT Registry #15385, covers 70,884 applications for 48 entry-level customer-service postings in the Philippines. The postings sat across 41 client accounts and 26 application sites in 19 cities, and required English fluency. Pay ran from Php 16,000 to Php 25,000 a month. Of the applications, 67,056 were eligible and randomly assigned between 7 March and 7 June 2025.

Compared with applicants interviewed by human recruiters, applicants interviewed by the AI voice agent were 12% more likely to receive a job offer, 18% more likely to start the job and 18% more likely to remain employed for at least one month. These are relative increases, not percentage-point changes. Transcript analysis found the AI interviews drew out more hiring-relevant information.

Among applicants who were given a choice, 78% picked the AI voice agent. In 5% of AI interviews, applicants ended the interview because they were unwilling to speak to an AI. In 7% of cases, the AI voice agent encountered technical difficulties. That 7% is relevant to your operations: a technical issue can interrupt an interview for a reason unrelated to the candidate's ability. Check connection recovery, device compatibility and what happens when a call drops.

This is useful evidence that an AI voice interview can work in a high-volume customer-service hiring process. But it is a working paper, not a peer-reviewed article, and it does not measure accent-specific recognition errors or score differences between language groups. It cannot establish that an AI interview will understand every applicant fairly. You still need to test speech recognition and scoring with your own candidate pool.

Do not confuse accent modification with better customer communication

Some support teams answer accent worries by changing the accent. Nicole Holliday of UC Berkeley and Melissa Baese-Berk of the University of Chicago tested that in a peer-reviewed study published 2 July 2026. The listeners were 202 American English speakers from the US Midwest, recruited through Prolific. They heard the four demo testimonials that the accent-modification company Sanas posts publicly, two speakers from India and two from the Philippines, each in an original and a modified version. Each listener typed what they heard in 40 excerpts and rated the voices.

It tested how people hear the voices, not how speech software does. Median transcription accuracy was above 80% for both versions. The apparent gain from modification disappeared once the speaker's country was taken into account: the overall difference was no longer significant (p = 0.082), while the difference by country was (p < 0.001), because Philippine speakers gained nothing. Their original speech was already transcribed at a very high rate.

Listeners rated the modified speech as more accented, harder to understand, and said they would be less satisfied with that agent, each at p < 0.001. Of 124 volunteered comments on how easy the voices were to understand, 56 called the original voices easy and 5 difficult. For the modified voices, 25 said easy and 27 said difficult.

The study used four demonstration recordings from one company, so it is not a verdict on every accent-modification product. For support hiring, the job-related question is whether a customer can understand the candidate and whether the candidate can listen, explain and respond appropriately. Do not use similarity to a preferred accent as a proxy for those skills.

Turn the research into a practical test for your next hiring pilot

Use these questions with any provider, including Fabric. Ask for clear answers before you use the platform to screen a full applicant batch.

If your pipeline starts with a phone screen, see our guide to AI phone screening for BPO hiring in India.

What Fabric lets you assess in a customer-support interview

For support roles, the aim is to assess whether a candidate can communicate clearly and handle a customer interaction, not whether they have a particular accent. Fabric's spoken-English assessment uses a 0–10 scale and does not score accent; it evaluates intelligibility, fluency, vocabulary, grammar and coherence, with an approximate CEFR level. Mother-tongue influence is described separately and is not scored. See how Fabric scores spoken English and what the score measures.

Fabric's customer-support interview uses customer scenarios and follow-up questions that respond to what the candidate says. That gives recruiters evidence of how someone handles a situation, rather than relying only on a prepared answer. The product page describes a 20–35 minute interview and a structured scorecard covering the role's skill areas. See the customer-support interview page for the workflow.

• Upfront disclosure. AI-use disclosure and recording consent are available in 35+ languages, so candidates can be told what the process involves before they begin.
• Technical failure. The interview opens in a phone browser without an app, is designed to work on weak 4G and can resume where the candidate left off after a dropped connection. Test this on the devices and networks your applicants actually use. For more detail, see the candidate-experience guidance for the candidate-side details.
• Language. Recruiters can select from 35+ languages and English varieties, including Indian, American, British, Australian and Singaporean English, as well as Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi and Tagalog. Do not assume that separate Hindi and English support means the system understands candidates switching between the two mid-sentence. Include that exact speech pattern in your pilot and ask how it is assessed.