What a 10-minute voice interview can and can't tell you

We build voice interviews, so this post covers what the modality is good for and the part of it we deliberately throw away. Voice reaches a depth a form never will. It also carries a temptation that has skewed hiring for as long as people have been interviewing people: judging how someone sounds instead of what they said. Our position is to name that temptation openly and then build the system so the machine cannot act on it.

What voice adds

A written application is a monologue the candidate rehearsed. A conversation has to be answered live, one question at a time. Voice does three things that a form and a one-way video recording cannot:

Adaptive follow-ups. The next question depends on the last answer. When a candidate says “we cut scope to ship on time,” the interview can skip the next item on a fixed list and ask the question that actually matters: “what did you drop, and what did that cost you?” A form asks everyone the same static questions, so it never gets to follow the thread, and the thread is usually where the real signal lives.

Spontaneous evidence. People rehearse their headline story, but nobody can rehearse for the third follow-up. Ask someone to go one level deeper than they prepared for and you get the unpolished specifics: the actual constraint, the tradeoff they made, the failure case and what they did about it. Those details are hard to fake in real time, which is what makes them worth scoring. A written answer shows you the version the candidate wanted you to see, and a conversation gets underneath it.

Lower friction. Talking is easier than writing, for most people and in most languages. There is no essay to draft, no form to grind through, no app to install and no account to create. The candidate opens one browser link whenever it suits them and is done in about ten minutes. That matters for completion, because the candidates who abandon a long written application are often the ones you most wanted to reach. Make it cheaper to show up and more people will.

What voice tempts you to judge

This is the uncomfortable half. The same audio that carries a candidate’s answer also carries their accent, their fluency, their pace, and the surface confidence of their delivery. Decades of research show that those signals move human interviewers, even though they say nothing about whether the person can do the job.

In a controlled study of interview judgments, an applicant’s accent and name shifted how favorably they were rated: an applicant with an ethnic-sounding name speaking with an accent was viewed less positively than the same profile without those cues, and those judgments fed straight into the hiring decision (Purkiss et al., 2006). In that study, the sound of a voice did work that only the content of the answer should do.

This is the trap built into the modality. The richest part of voice, hearing the person, is also the part that leaks bias. A tool that listens to a candidate and lets what it hears touch the score takes that old problem, automates it, and applies it to every candidate at once.

The line we draw

So we draw a hard line and enforce it in the architecture. Hure transcribes the conversation and scores only the words.

Concretely, that means no voiceprints, no tone or emotion analysis, no “confidence” detection, and no accent scoring anywhere in the system. The audio exists to be turned into text, and the text is what gets judged against the rubric. The interviewer keeps a record of what the candidate said and has no model of how they sounded; it was never built to have one.

On emotion, the law draws the same line. The EU AI Act prohibits AI that infers emotions in the workplace. Article 5(1)(f) takes workplace emotion recognition off the table entirely, and no compliance checklist makes it acceptable. Any interview tool that markets reading a candidate’s tone, stress, or “enthusiasm” is on the wrong side of it. We wrote up what that rule requires, and what NYC asks alongside it, in AI interviews and the law.

We can show the line held because every score cites a moment in the transcript. A rating based on how someone spoke, rather than what they said, would have no citable line behind it. Anchoring each score to the candidate’s words is what makes it auditable that sound never got a vote.

What 10 minutes is enough for

We should also say where the modality stops. Ten minutes is enough for a first round and too short for a full hiring loop, and we would rather undersell it than oversell it.

It is enough time to cover a sealed rubric’s competencies to evidence depth: ask about each one, follow the thread with a real question or two, and come away with a citable answer per competency. It is too short for a system-design deep-dive, a live coding session, or the long, winding conversation a final round is for. Stretching it to cover those would produce thin evidence dressed up as a thorough read, which is worse than admitting the boundary.

That boundary is deliberate, because it maps to the one stage where the math actually hurts: the first-round screening call. That is the interview you cannot afford to give every applicant by hand, and the one a consistent, evidence-backed conversation can replace without loss. We ran the numbers on why that stage is the one worth automating in the high-volume screening math. A 10-minute voice interview is built to replace the screening call, and the onsite stays as it is.

The close

Each part of the process has one job. The conversation draws out depth a form can’t reach, the transcript is the only thing that gets scored, and a person makes the call. The sound of a candidate’s voice is how you reach the substance of what they said, and the substance is all you should judge.