Voice Interview Reliability

Most voice interviews finish without a hitch: more than 95% complete end to end with no issues. This page covers the small remainder. What can go wrong, how to fix it, and what we do on our side. One thing before the list. Judge reliability across your pipeline, not on a single interview. If 100 candidates take a screen and a few hit a problem, that is normal for real-time voice AI, and a retake fixes it. If the same problem keeps appearing on one interviewer, that is usually a settings issue we can sort out with you quickly.

​

The interview ends before all questions are asked

Occasionally the AI decides the interview is finished when it isn’t, usually after a very detailed answer, a short ambiguous reply, or a long silence. This affects well under 5% of calls. The fix is a retake: open the candidate’s profile and click Allow Retake. Their earlier attempt won’t block the new one. Retake settings are covered in Configuring Your AI Interviewer. On our side, every early ending is detectable, and we run automated sweeps to catch them. We test each model and prompt change against a library of real calls where this happened. Because the behavior comes from the AI model itself, we can keep pushing the rate down but can’t promise zero. That is why retakes are free and take one click. If you see this twice on the same interviewer, check that the closing message and question list look normal, then send us the Interview IDs.

​

The AI talks over the candidate

Voice AI has to guess when someone has finished speaking. A long pause mid-sentence, background noise, or a quiet trailing word can read as the end of an answer. The most common cause we actually see is a setting: an interruption threshold below 3 seconds. Keep “Timing after interruption” at 3 seconds or more in the interviewer’s Advanced Voice Settings, and change one setting at a time from the defaults. All the settings are documented in Voice/Video Interviews. Candidates in a quiet room, speaking at a normal pace, rarely trigger this. If interruptions persist on default settings, contact support with the interviewer name and a couple of Interview IDs.

​

The call freezes, drops, or won’t start

Usually this is network or device trouble on the candidate’s end. Refreshing the page reconnects them; progress is saved and the interviewer picks up where things left off. Two setups cause most of the rest:

• Corporate-managed devices often have WebRTC disabled, which blocks browser calls entirely. Those candidates should use a personal device or take the interview by phone.
• Chrome is our most-tested browser and the most reliable choice for web interviews.

We run two independent voice providers, so an outage at one vendor degrades service rather than stopping it. During any incident we post live status updates, and we don’t charge for interviews that couldn’t start because of downtime on our side.

​

The voice sounds unnatural

Voice quality is subjective, and a voice that suits one role or region can sound off for another. If the interviewer sounds robotic, flat, or just wrong for your brand, try a different voice. You can preview every voice before assigning it, and swapping takes a minute: browse and add voices in Languages & Voices, then pick one under Identity in the interviewer’s Settings tab. A practical tip: run Test Now with two or three candidate voices and pick by ear. Voices also vary by language, so for multilingual interviews preview the voice in the language candidates will actually hear.

​

Pacing feels off or the pauses run long

If the interviewer takes too long to respond, leaves awkward silences, or feels rushed, that is almost always the timing settings rather than the AI itself. Everything is adjustable in the interviewer’s Advanced Voice Settings:

• “Timing before AI speaks” controls how long it waits after the candidate finishes before responding. Lower it if the gaps feel long, raise it if the AI jumps in too fast.
• “Silent seconds before prompt” controls how long a quiet candidate sits before the AI nudges them.
• “Timing after interruption” (the backoff) controls how long the AI waits before resuming after being interrupted.

Change one value at a time from the defaults and run Test Now between changes. The full list of settings is in Voice/Video Interviews.

​

The transcript doesn’t match what was said

Speech-to-text accuracy drops with background noise, weak microphones, and some accents. A headset in a quiet room prevents most of it. Answers with questionable audio are flagged for low speech clarity in the report, so they never quietly affect a score. Review flagged answers against the recording rather than the transcript. If a transcript clearly misrepresents a candidate, listen to the recording and offer a retake at your discretion.

​

Device setup

Candidates test their microphone (and camera, for video) before the interview starts, which prevents most connection problems. Encourage them to complete the check rather than click through it. You can also run Test Now on any interviewer to experience the exact candidate flow yourself. The full candidate journey is described in What Candidates Experience.

​

What the AI can’t promise

Voice interviews run on large language models, and a few behaviors are inherent to the technology rather than bugs with a permanent fix:

• Skipping or reordering a question, more likely on long question lists. We recommend 5–7 core questions.
• Hinting at what an answer should contain when a candidate asks for clarification, despite instructions not to.
• Rarely, misremembering or inventing a detail. This is the same hallucination behavior every AI product has.
• Imperfect handling of some accent, dialect, and language combinations.

We maintain a regression test suite built from real production calls, and every model or prompt change has to pass it before shipping. The rates keep going down. No vendor can honestly claim zero, so we would rather tell you where the edges are. Recordings, transcripts, and scores are all preserved so a person can always review the full context. Treat the AI interview as a screening signal and keep a human in the loop for final decisions.

​

What we commit to

On every plan: live status updates during outages, no charge for interviews that couldn’t start on our side, free retakes when an interview fails, low-clarity flagging in reports, and every reported edge case logged into our regression suite. Support depth depends on your plan and how widespread the issue is: To qualify for P1 or P2, the rate must be measured against a sample of at least 25 interviews, or the same issue must occur 3 times in a row. This keeps the priority levels tied to real patterns rather than a single unlucky call.

​

Getting help

Enterprise customers can reach their account manager directly. Everyone else, email support. Either way, include the Interview ID (from the candidate profile URL), a short description of what happened, and whether a retake was offered. One-off issues are almost always resolved with a retake. Patterns are almost always a configuration change. For setup and product questions, see the FAQ and Troubleshooting.