Interviewing
What Actually Improves the Transcript of Your Answer
Slow down only in specific places, and only after fixing the audio. HireVue names a microphone problem, or speech its system cannot understand clearly enough, among the reasons an answer is not scored by AI at all and is sent to a person instead. None of the speech-recognition research behind this answer compares speaking pace against sound quality. So fix the room and the microphone first, then slow slightly at names, acronyms and numbers, where one wrong word changes what you said.
The takeSpeak slower, enunciate, flatten the accent: the standard advice here is about sounding like somebody else, and that is the part worth refusing. A transcript is an instrument with a documented failure mode, not a verdict on how you speak. An hour spent on a wired headset, a quiet room and the pronunciation of three proper nouns aims at failures a vendor has actually written down, and it costs you nothing you would rather keep.
Where Olive fits
Open a role and see what the work shows
An Olive assessment is not an interview, and nothing drawn from how you sound enters the findings: you work a real assignment with an AI assistant, think aloud or type, and a person writes or edits the findings on what happened. You see every word the employer sees.
Assess your shortlistDoes Speaking Slowly Actually Help?
Nothing cited here shows that it does. A 2022 study ran two cloud recognizers over English recordings from first-language speakers across UK regions and 430 second-language speakers, and found word error rates significantly higher for the second-language group on both systems 3. Speech rate was a significant factor for Amazon Transcribe in the authors' regression. The recordings were read paragraphs rather than interview answers.
A significant factor is not the same thing as a usable instruction.
The bigger limit is what a change in delivery can reach at all. When researchers compared identical short phrases spoken by Black and by white speakers, the error rate was about twice as large for the Black speakers on each of five commercial systems as they stood in 2019 2. The authors traced that to the acoustic model rather than to grammar or word choice.
So a habit like reaching for more formal vocabulary aims at the wrong layer. The words in that comparison were already identical. What differed was the sound the model was working from. Accent is not automatically a wall either. Testing five systems on read sentences from speakers whose first languages were Arabic, Chinese, Hindi, Korean, Spanish or Vietnamese, Whisper and AssemblyAI reached mean match error rates of 0.054 and 0.056 7. The authors describe that as approaching human-level accuracy, on a corpus built for pronunciation research. The six first languages are all well resourced in recognition training data, and the paper is a preprint.
Read sentences in a clean recording are close to the easiest condition there is, and an interview answer is close to the hardest.
Whether your own round scores the transcript at all is a separate question, and what an AI interview actually scores covers how to find out.
Fix the Audio Before the Delivery
Start with the audio, because a vendor has written down when an answer stops being scored by AI at all. HireVue's 2025 explainability statement gives two conditions: the candidate consented to AI scoring, and the system detected enough content in the response. It names an answer that is too short, an answer not given in the allotted time, speech the system cannot understand clearly enough, and a microphone problem among the ways that fails 1.
When either condition is unmet, the statement says the response is not scored with AI and is flagged for human review instead. HireVue calls that a small proportion of candidates and publishes no number for it. Being routed to a person is not the same as being read favorably, and this is HireVue's process only; another platform may simply score a thin answer low.
It is the one route from your equipment to your result that a vendor has actually written down.
So spend the preparation there. Wired earbuds with a microphone sit closer to your mouth than a laptop lid does, and a small carpeted room returns less of your own voice than a kitchen. If the platform offers a practice recording, use it on the audio rather than on the answer, and play it back quietly. If you have to lean in to hear yourself, careful diction will not add signal the microphone never captured.
- A wired headset microphone, not the one built into the laptop
- A small room with soft surfaces, and the window shut
- Everything that pings set to silent, including the phone in the next room
- One test recording, played back before the first real answer
Hiring teams get the same problem described from the other chair, where a candidate's words arrive through a machine layer. What employers are told about judging an interview that ran through AI translation is to treat a garbled sentence as the pipe's error. Mark it unreadable, ask the question again, and judge what the candidate decided.
None of that advice asks anything of your voice.
Where Slowing Down Is Worth It
Slow at the words where a wrong transcription changes the meaning: your own name, the companies you worked for, product and tool names, acronyms, and numbers. A mangled adjective still leaves a readable sentence. A mangled employer name leaves a reviewer reading about a company you never worked for, and nothing in the surrounding sentence repairs it.
Say an acronym and then unpack it, once. "SOC 2, the security audit" costs about two seconds and gives both the recognizer and the reader a second chance at the term. The same goes for a number the story turns on.
No study here tested that habit; it is a judgment about consequences.
The rest of the answer does not need the same treatment. Slowing everything down flattens the delivery, and the recording can still be played back by a person. Write the three or four proper nouns you expect to say on a card before you start, somewhere you can see it. Saying a name you have just read is easier than retrieving one while a timer runs.
Deliberate pacing on four words is a small adjustment inside your own voice, and a different thing from trying to sound like a different speaker.
What Slowing Down Cannot Fix
Nothing in the research here shows delivery closing a recognition gap that tracks who is speaking. In a Dutch study, non-native speakers were grouped by their Common European Framework level, and read-speech word error rates came out at 57.2%, 56.5% and 57.0% across the three lowest levels 4. The authors write that, perhaps surprisingly, they do not see a reduction with an increase in level, on what is one Dutch recognizer with small groups at each level.
A machine transcript is not a measure of how well you speak a language.
The gap also carries a cost that arrives long before any interview. In a two-week diary study, 28 of 30 African American participants, all native English speakers, reported modifying their dialect so that voice technology would understand them, with examples such as talking "real clear" and dropping slang 6. Twenty-three of them reported frustration when errors occurred. Those are thirty paid participants describing consumer devices rather than a hiring platform, and the paper records what the accommodation cost the people doing it.
The vendor's own figures sit beside that research rather than replacing it. HireVue reports internal testing of Rev.ai, the speech-to-text supplier its scoring reads from, at a 7-10% word error range across White, Black, Hispanic and Asian applicants, against 15-30% for the other services it tested 5. That analysis is unpublished, gives no sample size or method, and reports a range rather than a figure per group.
If speech is a real barrier for you rather than nerves, asking about the format is the move with somewhere to go. Whether an AI interview can score you down for an accent or a disability covers what to ask, and how to request an accommodation on an AI interview covers how to send it.
A request is something a process can answer. A performance is not.
Common questions
Does speaking slowly lower my score?
None of the sources behind this answer says pace is scored. What pace changes is the transcript, and it changes it in one narrow place: slowing at a name or a number stops a wrong word from landing. Slowing everything mostly changes how you sound to whoever plays the recording back. On one published platform, HireVue's, an answer is not scored by AI when the system does not detect enough content. It names an answer that is too short, one not given in the allotted time, speech it cannot understand clearly enough, and a microphone problem, and it sends those responses to a person instead.
Should I repeat my name and my employers' names?
Once, said clearly and without hurrying, is usually enough. Those are the words where a wrong transcription changes the meaning of the sentence around them, and a reviewer reading about a company you never worked for has nothing in the transcript to catch the error with. Spell an unusual name if the format gives you the room. Repeating every proper noun twice costs you answer time you need for the substance.
Will better English lower the error rate?
The one study behind this answer measured Dutch, not English, and there higher proficiency did not lower the error rate. Read-speech word error rates for non-native speakers came out at 57.2%, 56.5% and 57.0% across the three lowest proficiency levels, and the authors report seeing no reduction as the level rose. That is one recognizer, one language, and small groups per level, and it says nothing about English systems. Treat it as a reason not to read a transcript as a verdict on your language, not as a reason to stop learning one.
What should I do if the transcript is visibly wrong?
Note what it got wrong and raise it with the recruiter in writing, if the platform shows you the transcript at all. HireVue's published process sends a response its system cannot understand clearly enough to human review, so a person listening is something that process already contains, though the statement describes no way for a candidate to request it. That is no guarantee of the answer you get. Keep the message short and specific: the question, the words that came out wrong, and what you actually said.
Is this the same as using an accent softener?
No. Slowing at a proper noun is a small adjustment inside your own voice, and an accent filter is another piece of software sitting between you and the platform. The recording is still there for a person to play afterwards, and none of the research behind this answer measures a filter against a hiring transcript. The pronunciation of three names is a smaller thing to work on, and it keeps your own voice.
References
- 1. 2025 Explainability Statement hirevue.com Supports the claim that a microphone problem or speech the system cannot understand clearly enough are named reasons an answer is not scored by AI and is flagged for human review instead.
- 2. Racial disparities in automated speech recognition pmc.ncbi.nlm.nih.gov Supports the claim that the error rate on identical phrases was about twice as large for Black speakers across five systems, which places the gap in the acoustic model rather than in word choice.
- 3. Language variation and algorithmic bias: understanding algorithmic bias in British English automatic speech recognition facctconference.org Supports the claim that second-language speakers had significantly higher word error rates on both recognizers, and that speech rate was a significant factor for Amazon Transcribe. The pool record states no direction for that factor, so the article states none.
- 4. Quantifying Bias in Automatic Speech Recognition arxiv.org Supports the claim that read-speech word error rates for non-native speakers did not fall as measured proficiency level rose.
- 5. 2025 Explainability Statement hirevue.com Supports the vendor's own unaudited internal figures for Rev.ai transcription accuracy by ethnicity, reported as a range rather than a per-group number.
- 6. "I don't Think These Devices are Very Culturally Sensitive." - Impact of Automated Speech Recognition Errors on African Americans frontiersin.org Supports the claim that 28 of 30 participants modified their dialect to be understood by voice technology, and that most reported frustration when errors occurred anyway.
- 7. Automatic Speech Recognition for Non-Native English: Accuracy and Disfluency Handling arxiv.org Supports the claim that on read sentences from non-native speakers two systems reached mean match error rates of 0.054 and 0.056, which the authors describe as approaching human-level accuracy.
7 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.