How accurate are AI interview transcripts? What the numbers mean and how to test yours

AI interview transcripts are accurate enough to search, skim and summarize, but not accurate enough to quote without checking. No vendor or independent study publishes an accuracy figure measured on job interviews as of November 2026. The closest evidence is research on similar audio: in OpenAI's own 2022 tests, the same model got 2.7% of words wrong on clean read speech, 13.8% to 17.6% on recorded phone conversations, and 36.4% on meetings picked up by a single distant microphone. The errors that matter in hiring, numbers, negatives, names and who said what, are the ones a headline accuracy figure hides.

This page explains what the numbers mean, why interviews score worse than benchmarks, and how to measure your own tool in about 15 minutes. For the specific failure where a model invents text, see AI note taker hallucinations.

How accurate are AI interview transcripts?

It depends far more on the audio than on the brand. The clearest public demonstration is in the paper that introduced OpenAI's Whisper speech recognition model, which tested one model on many kinds of English audio (Radford et al., "Robust Speech Recognition via Large-Scale Weak Supervision", OpenAI, 2022, Table 2, Whisper large-v2, word error rate after the paper's text normalizer):

Read this as a spread, not a score for any product. These are one model's 2022 results; the engines inside today's note takers differ and have improved, and benchmark transcripts follow their own spelling conventions (the paper notes that contractions in the Switchboard and CallHome references affect the scores). What carries over is the pattern: conversational speech, phone audio and distant microphones raise the error rate several times over compared with clean speech.

What does "95% accurate" actually mean?

Vendors usually mean a word error rate (WER) of about 5%. WER compares the machine transcript with a correct reference transcript and counts three kinds of mistakes:

WER = (substituted words + deleted words + inserted words) ÷ words in the correct transcript

Worked example (invented numbers). A two-minute answer contains 300 words. The transcript swaps 9 words for others ("fifty" for "fifteen"), drops 4 and adds 2 that were not said. WER = (9 + 4 + 2) ÷ 300 = 5%, or "95% accurate". If a 30-minute phone screen ran to 4,000 words at the same rate, the transcript would contain about 200 wrong words.

Here is what some vendors publish, as of November 2026. None of these pages gives the test audio or method behind the figure.

Even people are not perfect on conversational audio. Microsoft researchers measured professional transcribers at 5.9% WER on Switchboard and 11.3% on CallHome (Xiong et al., "Achieving Human Parity in Conversational Speech Recognition", 2016). A published percentage is a starting point for questions, not a substitute for testing on your own calls.

Why do interviews come out worse than the benchmarks?

• Laptop and phone audio. Built-in microphones pick up the room, and phone calls carry less of the voice than a headset does. Otter's own FAQ draws the same line between its notetaker's internal audio stream and a device microphone.
• Two people, sometimes at once. Interviews are conversation: interruptions, "mm-hm" over an answer, and quick back-and-forth, which is where the meeting and phone rows above sit.
• Who is speaking. A 2020 study of five commercial systems from Amazon, Apple, Google, IBM and Microsoft, run on recorded interviews, found an average WER of 0.35 for Black speakers against 0.19 for white speakers (Koenecke et al., PNAS, 2020). Accents and dialects the model has heard less often cost accuracy, and that makes transcript quality a fairness issue.
• Names, employers and jargon. Rare words are the ones a model guesses. A product name or a certification can come out as a common word that sounds alike.
• Pauses. A study of Whisper found roughly 1% of transcriptions contained whole phrases that were never said, more often for speakers with long silences (Koenecke et al., "Careless Whisper", ACM FAccT 2024). Interviews are full of thinking pauses.
• Language mixing. A candidate who switches language mid-answer can defeat a single-language setting; see multilingual interview transcription.

Which transcript errors matter in a hiring decision?

WER counts every word equally, so "uh" dropped and "haven't" heard as "have" cost the same. In a hiring record they are not the same. Four kinds of error change what a scorecard or submittal says: numbers (pay, notice period, team size, dates), negatives, proper names, and speaker labels that hand the interviewer's words to the candidate. A transcript with a 5% WER can be perfect on all four, or wrong on the one sentence that decides the scorecard. That is why a test should count those errors separately, and why the interview transcript example marks them in a raw transcript next to the corrected one.

How accurate are speaker labels?

Speaker labels come from one of two methods, and the method matters more than the model. Diarization guesses who is speaking from the voices in one mixed recording, and it fails most on short turns, crosstalk and similar voices; how it is measured is covered in what is speaker diarization. Channel separation records each side on its own stream, for example the interviewer's microphone and the call's audio, so the interviewer-or-candidate question is answered by where the sound came from. Channel separation does not help when several people share one stream, such as a panel in one room or both people on one microphone in person; panels are covered in how AI note takers tell speakers apart in a panel interview.

Interview Signal, which we make, uses channel separation: the microphone is labelled as the interviewer and the meeting's sound as the candidate, and any scorecard quote it cannot find in the transcript is removed. It still cannot fix a word the speech-to-text provider mishears, so the checks below apply to it too.

How do you measure your own tool's accuracy in 15 minutes?

1. Pick three two-minute excerpts from real interviews you have consent to use: one phone screen, one video call, and one with an accent, fast speech or heavy jargon.
2. Write the correct transcript of each by listening, replaying as often as needed. Do not edit the machine version; you will anchor on it.
3. Count substitutions, deletions and insertions against your correct version, ignoring filler words and punctuation, and work out WER with the formula above. NIST's free SCTK scoring toolkit does the alignment and counting if you prefer software.
4. Count critical errors separately: every wrong number, flipped negative, misspelled name and wrong speaker label.
5. Compare tools on the same excerpts. Running two tools on the same audio is the only fair comparison; a vendor's figure from different audio is not.
6. Repeat when something changes: a new tool version, a new language, a new interview format.

A tool that scores well overall but makes critical errors in every excerpt needs a checking step before anything reaches a scorecard. The checks in AI note taker hallucinations take about ten minutes per interview.

How do you get more accurate interview transcripts?

• Wear a headset or use a directional microphone, and ask the candidate to use headphones so your voice does not echo through their microphone.
• Prefer a capture method that takes the meeting audio directly over a microphone listening to speakers.
• Take phone screens through the computer or a calling app rather than holding a phone up to a laptop; see how to transcribe a job interview for each format, and does an AI notetaker capture both sides of a phone interview for what each capture method hears.
• Add employer names, products and certifications to the tool's custom vocabulary if it has one; Otter's FAQ recommends this for specialized terms.
• Ask the candidate to spell their name and any unusual employer at the start, and confirm pay figures and dates in writing afterward.
• Avoid talking over answers. A quiet "mm-hm" in the middle of a sentence is a common source of misattributed words.

Is an AI transcript accurate enough to base a hiring decision on?

On its own, no. Use it as a searchable memory of the conversation and as the source of quotes, then confirm each quote, number and negative that goes into a scorecard, a submittal or the ATS. Keep the corrected version as the record, and if a candidate disputes what they said, go back to the audio, or to your own notes if no audio was kept. Because accuracy varies by speaker, a process that leans on unchecked transcripts can disadvantage some candidates more than others, which is a good reason to make the check routine rather than occasional. For comparing tools on accuracy and everything else, see interview transcription software for recruiters.