How AI evaluates interview transcripts is a two-step process: speech-to-text software first turns your spoken answer into written text, and then a language model reads that text and rates it against the rubric the employer set for each competency. Whatever the transcript doesn't capture, such as a misheard number or a point you only gestured at, usually can't earn credit.
That makes the transcript the version of you that gets scored. This guide explains both steps using HireVue's published description of its system and other vendors' documentation, checked in October 2026, then shows the same answer as a messy and a clean transcript. Take the assessment yourself and follow the employer's rules — use CoPilot Interview to practice beforehand.
How AI evaluates interview transcripts: the pipeline
| Step | What happens | What it keeps | What it loses |
|---|---|---|---|
| 1. Speech-to-text | Your audio is converted to words | Your words, in order | Tone, emphasis, pauses, gestures, and anything on a screen or whiteboard you didn't say out loud |
| 2. Language understanding | A model reads the text for meaning | What you claimed, did and achieved | Words that were misheard or never made it into the text |
| 3. Rubric scoring | That meaning is compared with a rating scale for each competency or question | A rating, and sometimes an explanation | Content unrelated to what the question targets |
HireVue's 2025 Explainability Statement, still published as of October 2026, describes exactly these three stages for its AI-scored video interviews: transcribing spoken words to text, understanding the meaning of the text, then scoring the answer (HireVue Explainability Statement). It also says its AI "relies only on what is said by the candidate," with no facial, body-language or tone-of-voice analysis. Other vendors may work differently, but transcript-first scoring is the common pattern.
Step 1: speech-to-text turns your answer into text
HireVue says it uses a third-party speech-to-text system from Rev.ai, trained on human-transcribed audio across many topics, dialects and accents, and that it tested transcription accuracy for native and non-native English speakers before adopting it. Mercor says its AI interviewer generates a transcript of your responses, and that it can't share interview recordings or transcripts with candidates (Mercor AI interview support).
Even good speech recognition makes mistakes, and interviews are full of the words it finds hardest:
- Technical terms and product names, especially ones that sound like ordinary words.
- Acronyms, which can come out as a different word entirely.
- Similar-sounding numbers, such as fifteen and fifty.
- Names of people, companies and places.
- Words lost to noise: a fan, a second voice, a laptop microphone across the room.
- The end of your answer, if the timer runs out or a voice bot decides you've finished.
Step 2: a language model matches the text to a rubric
Once your answer is text, a model reads it for meaning. HireVue says its language model is fine-tuned separately for each competency and judges what an answer means "regardless of the specific vocabulary used." The result is scored against a behaviorally anchored rating scale with five levels, from novice to expert, built from expert human ratings of real interview answers. For its conversational AI Interviewer, HireVue says each response is evaluated against role-specific criteria, marked as meeting them or not, and explained (HireVue AI Interviewer).
What that means in practice: the model needs claims it can match to the rubric. A sentence like "I rebuilt the report so it ran in two hours instead of fifteen" gives it a clear action and result. Fragments, half-finished sentences and restarts that never resolve give it less to work with. Keyword lists don't help either, because the model is reading meaning, not counting terms. To see what the top and bottom of a rating scale describe, read our explainer on behaviorally anchored rating scales.
When the transcript is too thin to score
HireVue says its AI text scoring needs two things: your consent, and enough content in the response. It lists three reasons an answer might fall short: it was too short or not finished in the allotted time, it wasn't spoken clearly enough to be understood, or a microphone problem got in the way. Those answers are not scored by AI and are flagged for human review instead.
Human review isn't a penalty, but it means your answer wasn't evaluated the way the employer set up the process, and a thin answer is thin for a person too. The fix is simple: test your microphone before you start and give complete answers rather than one-liners.
Messy vs clean: the same answer as a transcript
Question: "Tell me about a time you improved a slow process."
Messy transcript (example): "um so yeah at my last, well, the one before, we had this, the, uh, reporting thing was really slow and so we kind of looked at it and, you know, I think it was like fifty, no, fifteen hours or so, and we fixed it with the, uh, sequel stuff and it was way better after."
What the scoring has to work with: no clear situation, "we" hiding your role, a number that appears twice (the transcript may keep "fifty"), "SQL" heard as "sequel" with nothing around it to explain it, and a result of "way better."
Clean transcript (example): "At my previous job as a financial analyst, our month-end sales report took about fifteen hours to build by hand. I rewrote the three slowest steps as SQL database queries and tested them against two months of old reports to make sure the totals matched. The report now takes about two hours, and I documented the queries so two other analysts can run it."
Same story, same person. The clean version names the situation in one sentence, uses "I" for the actions, puts the acronym next to the plain word "database" so the meaning survives even if it's mistranscribed, and ends with a measured result. A few "um"s are normal and are unlikely to matter on their own; for more on that, see whether filler words hurt AI interview scores.
Transcription errors and how to reduce them
| Problem | What the transcript might show (example) | Fix |
|---|---|---|
| Acronyms and jargon | "sequel" for SQL, "cooper netties" for Kubernetes | Say the term once with a plain-word anchor: "Kubernetes, the container platform." |
| Similar numbers | "fifty" when you said "fifteen" | Give numbers a unit and a comparison: "fifteen hours, down to two." |
| Self-corrections | Both versions appear in the text | Pause, then restate the whole sentence cleanly. |
| Talking over the interviewer | Overlapping or missing words | Let the question finish; ask for it to be repeated if you need to. |
| Noise or a distant microphone | Gaps and wrong words | Use a headset or wired microphone in a quiet room and test it first. |
| Fast speech that trails off | Dropped words at the ends of sentences | Slow down slightly and finish every sentence. |
| Running out of time | The result never appears | Lead with the answer and reach the result well before the limit. |
If you speak English as a second language or have a strong regional accent, the same fixes help most: a steady pace, complete sentences and key names said clearly once. Our AI interview tips for non-native English speakers go further, and how to structure answers for AI transcription covers signposting and pacing.
Practice: read your own transcript
- Record one answer with your phone's voice typing or any captioning tool, so you get text, not just audio.
- Read the text cold. Could a stranger find the situation, your actions and the result? Circle every misheard word.
- Rephrase what broke. Add a plain-word anchor to each misheard term and restate numbers with units.
- Get a scored read. Our free AI interview assistant includes a Mock Interview mode that reads questions aloud, listens to your spoken answer and scores it out of 10 with what to fix and a model answer, then gives you a report card. The free plan includes 15 minutes every day, with no card and no signup.
- Repeat with nothing installed using the browser AI mock interview.
For the full setup checklist, including microphone and room, see how to pass an AI interview.
Find out what your transcript is missing
Answer out loud, get a score out of 10 and a note on what to fix, then try the answer again in your own words.
Try a free AI mock interviewFAQ
How does AI evaluate an interview transcript?
Speech-to-text software first converts your spoken answer into text. A language model then reads that text for meaning and compares it with the rating scale or criteria the employer set for each competency or question, producing a rating and sometimes a short explanation.
Does the AI hear my tone of voice?
It depends on the vendor. HireVue says its AI-scored interviews rely only on what you say and do not analyze tone of voice, facial expressions or body language. Other tools may differ, so read the notice before your interview.
What if speech-to-text mishears me?
You usually cannot edit the transcript, so prevention matters: use a good microphone in a quiet room, explain acronyms with a plain word, and restate numbers with units. If a technical problem affected your interview, tell the recruiter promptly.
Do filler words show up in the transcript?
Some systems keep them and some drop them. A few fillers rarely change the meaning of an answer, but frequent fillers and restarts make the text harder to follow for both the model and any person who reads it later.
Can I get a copy of my AI interview transcript?
It varies. Mercor, for example, says it cannot share interview recordings or transcripts with candidates. Elsewhere, ask the employer. Privacy laws in some places give you rights over your personal data, so check the law where you live.
Does speaking with an accent lower my score?
HireVue says it tested its transcription system for native and non-native English speakers before adopting it. Accuracy can still vary, so speak at a steady pace in complete sentences, and ask about accommodations if the format puts you at a disadvantage.