What Happens to Your Ums Between the Recording and the Score

What Does the Score Actually Read?

On HireVue, the one large platform that publishes an answer, the score is built out of words. Its 2025 explainability statement says the AI relies only on what the candidate says, using no video analysis and no other audio characteristics ¹. Facial expressions, body language, background and tone of voice are named in that document as things it does not assess ¹.

HireVue is describing its own product there rather than reporting an independent audit, and the statement is scoped to its own AI-scored interviews. It says nothing about the smaller platform whose name you had never heard before the invitation arrived, and nothing about what a recruiter notices watching the same video afterwards.

The recording does not vanish because a model ignored it.

The statement leaves one thing about an um open. The pipeline runs speech to text first, then a language model, then a score benchmarked against expert human ratings of answers to similar competency questions ¹. Whether the transcriber writes a filler down as a word or drops it is not published there. So an um could in principle reach the model as text, even though the sound of it does not.

The statement gives a candidate no way to check.

Some transcribers have failure modes of their own, and one audited model stumbled on long silences. An audit of OpenAI's Whisper found that roughly 1% of transcription segments contained phrases appearing nowhere in the underlying audio, and that these fell disproportionately on speakers with longer non-vocal stretches ². That was one model version tested against an aphasia research corpus rather than against interview recordings, and four other commercial services produced no comparable inventions on the same segments.

On that one model, a long pause was less a hesitation than a gap the transcriber had to cross.

Why Do Practice Tools Grade Your Fillers?

Because pace and filler words are what a microphone can count. Yoodli's interview practice page says the tool generates a speaking report with analytics such as pacing and filler words, plus suggestions to improve ³. Big Interview's VideoAI page says it evaluates pace of speech, vocabulary, eye contact and body language, alongside cutting filler words ⁴.

Pace, eye contact and body language are delivery signals that HireVue states its own scoring does not use ¹⁴. Both pages are vendor marketing with no method, benchmark or accuracy figure attached to any of it, and neither company is the one running the assessment you were actually sent.

Independent research points mostly the same way, with one exception worth carrying. Hickman and colleagues decomposed the weights inside automated video interview models. For four of the five Big Five traits, the models weighted verbal behaviour, meaning the words themselves, above voice and above face and body ⁵.

Extraversion was the exception. On average its models put 38% of the weight on paraverbal cues such as loudness and speech rate, 34% on the words and 28% on face and body ⁵. The shares also moved substantially between samples, which is the finding rather than a footnote: which cues matter depends on the sample a model learned from.

Those were academic models rather than any vendor's product, and the thing being predicted was how raters perceived students and crowd workers in mock interviews, not who got hired ⁵. What it supports is narrow. On a tool that publishes nothing about itself, delivery cannot be ruled out; on HireVue, which does publish, the vendor rules it out of the AI score in writing.

What Does a Human Reviewer Hear?

Everything the model skipped. The video still exists, someone on the hiring side may watch some of it, and a person hearing four false starts in the opening sentence may write that down as nerves. That impression comes from a human rather than from a scoring model, and it is the part of the round genuinely about delivery.

Hiring teams are warned against putting exactly that impression in writing. The employer-side piece on whether an AI notetaker's summary can go in the hiring file argues that a line like "seemed nervous and rehearsed" is a conclusion about a person. It is not a record of anything they said, and it should be struck before the file closes.

Another employer-side piece asks hiring teams whether the async video round is worth keeping once scripted answers leave it separating candidates on delivery and barely at all on substance. Reading what the other side of the table is advised beats guessing at it from this one.

Smooth delivery is not the fix.

A reviewer who hears a stumble and then a specific number still has the number. The clean version of the same answer, with the specifics smoothed out, leaves them less to carry into the debrief. For the wider version of this question, what an AI interview scores and what it does not covers the rest of the format.

Check the Retake Rules Before the Fourth Take

Retakes are a setting the employer chooses rather than a property of the product. HireVue's statement lists interview question retries as a Yes or No configuration, with unlimited as the recommended setting ⁶. It says plainly that how a candidate experiences the assessment is configurable by the company deploying it ⁶. Read the instructions screen instead of assuming one shot.

Once an answer is submitted, that door closes. HireVue's candidate help center says you cannot return to a submitted question, whether the interview is still running or already finished, and that only the hiring company can authorize a new attempt ⁷.

There is a floor under a genuinely broken take. The statement says a response the system cannot score is not scored by the AI and goes to human review instead ⁸. An answer too short, one not given in the allotted time, speech that cannot be made out, and a microphone problem are the ways it names ⁸. The vendor calls that a small proportion of candidates without publishing a figure, and going to a person says nothing about the outcome.

So take the pause.

Start the sentence again if it came out wrong. Say the number you were reaching for, name the part that went badly, and stop when the answer is finished. The take that can cost you something is the fourth or fifth, where the delivery has improved and the detail has quietly gone.

See how it works >