When the Transcript Gets It Wrong, the Score Reads the Mistake

Does a Wrong Transcript Change What Gets Scored?

Yes, on the two platforms that have published what their scoring reads. HireVue's 2025 explainability statement describes a three-stage pipeline: speech to text, then a language model, then a score ¹. The same document says the AI uses no video analysis, and does not assess facial expressions, body language or tone of voice ¹. Vervoe's candidate help center says the same about audio and video answers: the AI reviews only the transcript ².

Neither document is an audit; both are vendors describing their own products.

Take them at face value on that narrow point and the consequence follows. If the words are the input, the words the recognizer produced are what the model read. A dropped clause is a shorter answer. A substituted word is a different answer.

Neither shows up in the score as an error, because to the model it was never an error at all.

A person on the employer's side has something wider to look at. Vervoe publishes the half of this that helps you: the video and audio are transcribed and provided to the employer along with the recording ². So the mistake the model made is recoverable by a human reader in a way it is not recoverable by the model, and that asymmetry is the whole reason to say something.

The recording is the evidence. The transcript is only a reading of it.

How Often Does Recognition Actually Fail?

Often enough to be measured, and not evenly. Koenecke and colleagues tested five commercial speech recognizers in 2019 on recorded interview audio ³. Using the authors' own threshold for an unusable transcript, 23% of snippets from Black speakers came back unusable, against 1.6% of snippets from white speakers ³. That audio was sociolinguistic interview speech rather than answers spoken into a laptop, and no current interview product was in the test.

HireVue publishes its own numbers beside that. It names Rev.ai as its speech-to-text supplier, and reports internal testing at a 7-10% word error rate range across White, Black, Hispanic and Asian applicants ⁴. It puts human transcription at roughly 5-10% for comparison ⁴.

Those numbers come from an internal analysis HireVue has not published, with no sample size, no audio corpus and no per-group breakdown. They sit beside the research without overturning it, and they measure transcription only ⁴.

The other failure worth naming is invention rather than omission. In a 2024 study of one general-purpose speech model, roughly 1.4% of transcriptions in the test set contained a hallucination ⁵. Of those hallucinated transcriptions, 38% carried at least one identified harm, including made-up names ⁵. That model has not been shown to be the one any interview vendor uses, and the audio was clinical speech rather than an interview answer. The runs also date from spring 2023, and the authors report that later updates resolved many of the specific cases.

An invented sentence reads as fluently as a real one, which is what makes it hard to spot.

An average error rate tells you nothing about your answer.

Employers reading about the same problem are told something specific. Keep the audio, and judge what the candidate said rather than how cleanly it came back. That is the argument in how accurate AI interview transcription is, and whose words it gets wrong.

Write to the Recruiter the Same Day

Send a short factual message while the process is still open, because the people who can act on it sit on the employer's side. Name the question, say in one line what you actually said, and ask whether a person can compare the recording with the transcript. Keep it to facts about the recording rather than an apology for your delivery.

• the question number, or its first few words, so nobody has to guess which answer you mean
• one sentence naming the phrase that came out wrong and the phrase you said
• the request: that a person read the transcript against the recording before the file moves on
• the date and rough time you recorded, so the right recording is easy to find

Send it from the address you applied with.

Specificity matters more than tone. A message naming one question, one phrase and one request reads as a bug report, and a recruiter can forward it in a second. A message about how badly the whole thing went cannot be forwarded anywhere. Attach nothing, and let the recruiter ask for more.

Two published constraints shape what you can reasonably ask for. HireVue's help center says a submitted answer cannot be reopened, and that only the hiring company can authorize a retake, which it does by sending a new interview link ⁶. The vendor says it is not involved in hiring decisions and cannot grant one itself ⁶.

Whether you could have re-recorded before submitting is a separate setting. HireVue lists unlimited question retries as its recommended configuration while stating plainly that the company deploying the assessment decides how the candidate experiences it ⁷.

The second constraint is access. After a HireVue interview a candidate has no direct way to review the recording through the platform ⁸. The help center points anyone who wants to see it back to the hiring company, which controls the recordings and decides whether to release them ⁸.

So on HireVue, the ask goes to the employer.

Whether anyone opens the recording at all is its own question, and whether a human ever watches an AI-scored interview sets out what the evidence supports.

What If the Interview Already Closed?

Send the message anyway, and keep it factual. A rejection does not close the record. The recording and the transcript sit with the employer for as long as it keeps them, and a note naming a specific recognition error is a different object from a note asking to be reconsidered. It may change nothing about this role.

It costs one message.

A floor sits under this on at least one platform. HireVue says an answer it cannot understand clearly enough, or one cut short by a microphone problem, is not AI-scored at all and is flagged for human review instead ⁹.

An answer from a candidate who declined AI scoring lands in the same queue, since consent is the other condition HireVue names ⁹.

Being routed to a person moves the response out of the model's hands and says nothing about the outcome ⁹.

The vendor calls this a small proportion of candidates in any cohort and publishes no figure for it ⁹. It does say the rate is monitored for each customer, so somebody on the employer's side can in principle be told how often it happened ⁹. None of that transfers to a platform which has said nothing in public about how it handles an unclear answer. On another system a thin transcript may simply score low.

Two things make the note worth sending even after a no. It records that the error happened, close to when it happened. And an employer running many of these has its own reason to know when the recognizer failed, separate from anything about your application.

None of this relitigates a hiring decision. It puts one answerable question in front of the only people holding both the recording and the transcript: whether the words on the page are the words you said.

Read the evidence >