Can AI interviews understand Indian English and mixed-language answers?

Updated 3 November 2026

If you hire customer-facing, sales or support roles across India, your AI interviewer needs to understand how candidates actually speak. A transcript that changes a client name, a tool, a number or a qualification can leave recruiters with the wrong impression of an answer. One 2023 Indian-English benchmark reported word error rates between 20.7% and 30.0% for the commercial systems it tested. That is a reason to test carefully, not a prediction of how today's interview products will perform.

There is no single accuracy number that applies to every candidate or interview. Results change with the speech model, language, regional speech patterns, microphone and recording conditions. The useful question for a hiring team is whether the system can accurately capture the answers your candidates give, on the devices and in the languages they use.

What a transcription error can change in your hiring decision

Word error rate (WER) is a common way to measure transcription accuracy. It counts substitutions, deletions and insertions, then divides that total by the number of words in the reference transcript. If a 200-word answer has a WER of 25%, that means the system made 50 word-level errors under this measure. It does not mean that exactly 50% of the answer's meaning was lost: some errors matter far more than others.

For a recruiter, a wrong name, number, tool or qualification can matter more than several small grammatical errors. The interactive example below shows why the type of error matters, not just the overall percentage.

The Svarah benchmark also tested everyday-use prompts containing brand names, bank names, food items and document IDs. Whisper Large scored 7.2% WER across the full benchmark and 11.2% on these use-case prompts. For your hiring process, include the words recruiters rely on: client names, software, certifications, years of experience, targets and numerical results.

A transcript can look complete while still changing a key detail. If a recruiter relies on it, that error may affect the review. Ask to see the original recording alongside the transcript and scorecard, and find out which parts of the audio and transcript contribute to the assessment. Do not assume every transcription error changes a score, or that every interview score is based only on a transcript.

Transcription and evaluation are separate steps. A system may capture the words accurately but assess the answer poorly; it may also mis-transcribe a good answer before evaluation begins. This page focuses on whether the words are captured correctly. For the separate question of how spoken communication is assessed, see how Fabric scores spoken English for BPO hiring.

What Indian-English research can and cannot tell you

One useful benchmark is Svarah, developed by AI4Bharat researchers and published at INTERSPEECH 2023. It contains 9.6 hours of English from 117 speakers across 65 districts in India. The recordings include read speech, open conversation and everyday-use prompts, collected through an Android recording app. The authors evaluated six open-source systems and two commercial cloud services, from Google and Microsoft, each in an India and a US setting.

This benchmark is from 2023, and commercial systems may have changed since then. Svarah covers read, extempore and everyday-use speech, not live recruitment calls. Its native-English reference scores come from separate datasets, so they are not direct comparisons.

The best result in the Svarah study came from Whisper large, an open-source system, with a 7.2% word error rate on that benchmark. That is evidence that Indian-accented English can be transcribed well under the right conditions. It is not proof that every model, product or recording setup will achieve the same result. Training data, model version and audio quality all matter.

Hindi-English mixed speech has been studied separately. The 2021 MUCS challenge, published at INTERSPEECH 2021, reported results on a realigned Hindi-English blind test set. The baseline systems were trained on 89.86 hours of Hindi-English speech, and they are research baselines, not products you can buy.

Do not compare these percentages as if they came from one controlled test. The studies use different speakers, datasets and systems. They cannot establish that Hinglish is a fixed percentage harder to transcribe than English. The MUCS figures below are challenge baselines on one Hindi-English blind test set, not a current benchmark of commercial tools. The practical lesson is simple: test mixed-language speech separately. A system that supports Hindi and English individually may still struggle when a candidate moves between them in one answer.

Why performance changes between speakers and settings

Why the same system can hear two candidates differently

Too little Indian speech in the training data. The Svarah authors point out that common English benchmarks such as LibriSpeech and Switchboard do not include speakers living in India. The amount and variety of Indian speech in training and evaluation data can affect how well a system handles Indian-English speech.

India has many regional language backgrounds and speech patterns. Svarah reported weaker results for some speakers whose first languages were less represented in available datasets, including Assamese, Bodo and Nepali. Do not judge performance from a single demonstration accent or one city. Include the regions and language backgrounds represented in your applicant pool.

Two languages in one sentence. In everyday conversation, a candidate may move between Hindi and English within a sentence, especially when naming tools, processes or technical terms. A system designed and tested only for one language at a time may not handle that naturally.

Real audio is often less controlled than test audio. Phone calls and browser interviews introduce different conditions: microphone quality, background noise, compression, echo and unstable connectivity. A model tested on clean audio may perform differently on a low-cost handset or a call taken in a busy room. Test the channel you intend to use. For more on phone conditions, see our guide to AI phone screening for BPO hiring.

A 2024 study of 262,578 public tweets from 16,710 Twitter accounts in Delhi and Mumbai found that the share of mixed Hindi-English text rose from about 42% in 2015 to about 60% in 2020, and stayed near that level afterwards. The study, published in 2024, concerns written social-media language in two cities, not spoken interviews or the whole of India. It cannot tell us how often your applicants use Hinglish, but it does show why mixed-language communication should not be dismissed as an unusual edge case.

If your candidates naturally mix languages, include that speech in your test. Do not ask them to read a prepared sentence. Use consented, natural answers to role-relevant questions and compare what the candidate said with what the system transcribed.

What to check when a product says it supports Hindi or Hinglish

Language support can mean several different things. Check which of these the interview actually handles:

Decide whether the role requires English or multilingual communication

The right language policy depends on the job. For an international voice process or a client-facing role where English is essential, assess spoken English against clear, job-related criteria. For domestic sales, collections, field operations or regional customer support, candidates may reasonably use more than one language. An English-only assessment in those roles could measure language choice rather than the communication skills the job requires.

Fabric lets the recruiter select the interview language. Its BPO offering also describes spoken-English assessment for roles where English is required. Match the selected language and scoring criteria to the job. For details on Fabric's BPO offering, see the Fabric BPO page.

How to test an AI interviewer with your own recordings

A published benchmark cannot tell you how a product will perform on your applicants' audio. Test the same set of consented recordings on the interview system you plan to use, including the devices, accents, languages and conditions found in your applicant pool.

Have a human transcriber create a reference transcript. Compare the system's transcript with it, counting overall word error rate and errors in names, numbers, tools and job-specific terms separately. Check the same answer in English, Hindi and natural mixed speech. During a pilot, track interview completion, transcription errors and false rejections. Have recruiters review a sample of rejected candidates to see whether the system screened out someone they would have progressed.

What to check in Fabric's language and scoring approach

For BPO phone screens and AI interviews, Fabric's current product material lists 16 languages plus five English variants, with the recruiter selecting the interview language. If candidates switch between Hindi and English within one answer, test that exact pattern in a live demonstration. Support for Hindi and English separately does not by itself establish how well the system handles mixed speech.

Fabric says accent itself does not change the spoken-English score; mother-tongue influence is described separately rather than scored. For English-language roles, it provides a spoken-English score from 0 to 10 and an approximate CEFR level. Read the explanation in how Fabric scores spoken English for BPO hiring.

Fabric has not published a word error rate for Indian-accented audio. We cannot therefore claim that its transcription is more accurate than the systems in the benchmark or assign it an accuracy percentage. Ask to see the transcript and scorecard together, then run a controlled pilot using consented sample audio.

Frequently asked questions

Can AI interview software understand Indian accents?

Some systems recognise Indian English, but performance varies by model, accent, vocabulary and recording conditions. The 2023 Svarah benchmark found a wide range of results across the systems it tested. It is not a current comparison of interview products. Test the product you plan to use with speakers, devices and vocabulary that reflect your applicant pool.

Can AI interviews handle Hinglish, where candidates mix Hindi and English in one sentence?

Some systems can process mixed Hindi-English speech, but the phrase can mean transcription, question delivery, follow-up conversation or scoring. Ask which of these functions work in mixed speech and test them using natural answers. Research benchmarks use different datasets, so there is no single Hinglish error rate that applies to every product.

Does a candidate get a lower score for mixing Hindi and English?

A candidate should not receive a lower score simply for mixing languages, unless the role genuinely requires a particular language and that requirement is assessed consistently. Check how language choice affects scoring, and compare the same answer delivered in different languages. The score should reflect job-related communication criteria, not an unspoken preference for one way of speaking.