Can a Voice Interview Tell It Isn't You Speaking?

Some remote interviews now analyse your voice as evidence, not just as speech to transcribe. The checks fall into three separate layers, and they are not equally reliable: matching your voice against an earlier sample is the oldest and steadiest of them, detecting a synthetic or cloned voice is the newest and the one that degrades most on ordinary call audio. As of 21 October 2026, the strongest accuracy figures in this space are vendor self-reports on proprietary datasets, while the published independent evaluations show error rates an order of magnitude higher once the audio has been through a codec or a phone line.

That gap is the practical problem. It means an honest candidate on a compressed connection sits closer to the flagging threshold than the marketing numbers suggest. This article separates the three layers, states what each one can and cannot establish, and sets out what to ask for before a voice interview. It is a reading of vendor documentation and published research, not legal advice.

Three different things listen to your voice

Candidates tend to treat "the AI is analysing my voice" as one capability. It is three, with different purposes and different failure modes.

Transcription converts your speech to text for reviewers to read, search, and quote. It makes no claim about who you are. It is also the layer most likely to cause you quiet harm, because errors in a transcript follow you to whoever reads it later — the evidence on how speech recognition handles accented speech in interviews covers that separately.

Speaker verification compares the voice on this call against an enrolled voiceprint or an earlier interview stage, and answers "is this the same person." This is mature technology with decades of deployment in banking telephony.

Synthetic-speech and bot detection asks a different question: "was this audio produced by a human vocal tract or by a machine." It does not need a reference sample of you. It looks for artefacts of generation, and it is the layer where the accuracy claims and the independent evidence diverge most.

A tool can run any one of these without the others. An interview that transcribes you is not verifying you. A tool that flags a cloned voice is not confirming your identity — it is only asserting that the audio looks generated.

What the vendors say they do

Pindrop is the most explicit vendor in this category, and its hiring-facing product is the clearest published example of voice checks written into a candidate record.

Pindrop's Pulse for Meetings page describes detection of "proxy candidates and identity swaps across interview stages," integration with Zoom, Webex, and Microsoft Teams, and three check categories: video and audio liveness, a passive comparison of the speaker's voice against an enrolled voiceprint, and location risk including VPN use. The page states that results write back to the candidate record automatically to give recruiters an audit trail, and cites "2,300+ detection signals" without describing them (Pindrop, Pulse for Meetings). It names integration with "leading hiring platforms" but does not identify them.

The accuracy figures on that page are 99% detection for known deepfake engines, over 90% for new or unseen engines, and a false positive rate under 1%, derived from a proprietary dataset of 20 million samples across more than 120 voice generation systems. Those are vendor-reported results on vendor-held data, and the page attaches them to the audio deepfake product rather than separately to interview video conferences. It lists no limitations or conditions under which detection is weaker.

Separately, on 16 September 2026 Pindrop announced BotStopper, which it describes as identifying whether a caller is human or machine in about two seconds, built on the same Pulse engine and drawing on an "AI Voice Consortium" registry of more than 5,000 AI voices (Pindrop press release, 16 September 2026). That release is aimed at contact centres rather than hiring, but the registry approach matters to candidates for a specific reason: recognising a known commercial AI voice is a much easier problem than recognising synthesis in general, and a system's headline accuracy may be carried largely by the former.

What independent evaluation shows

Two published sources put numbers on the harder problem, and both are more pessimistic than any vendor page.

The ASVspoof 5 challenge is the field's standard open evaluation, run by researchers at the National Institute of Informatics, EURECOM, the University of Eastern Finland, and others, with co-authors from industry including Pindrop. Its evaluation paper reports that in a cross-dataset test, the four top-performing open-condition systems had equal error rates above 10% on most out-of-domain sets. The authors describe generalisation across attacks and datasets as remaining a major unsolved problem — the field's "holy grail" (ASVspoof 5 evaluation paper, arXiv 2601.03944, v3 dated 11 April 2026). The same evaluation found neural codecs among the most difficult conditions, and noted that several top systems were poorly calibrated, meaning their scores do not translate cleanly into a single reliable decision threshold.

A deployment-focused report from Brno University of Technology and the voice-biometrics firm Phonexia is more direct about what happens to these systems in production. Published as a preprint on 18 August 2026, it reports that a detector trained without representative codec augmentation saw its miss rate rise from roughly 4% to 60% after a single, non-aggressive codec pass. In narrowband telephone conditions, false-alarm rates rose from roughly 5% to 70%, with equal error rates of about 16% under AMR-NB and about 25% under G.711 (The Last Mile of Deepfake Speech Detection, arXiv 2608.17585). The authors also note that per-attack error rates can each look acceptable at individually tuned thresholds while a single deployed threshold performs noticeably worse, and they caution that their own figures are illustrative internal results rather than reproducible benchmarks.

The direction of those numbers is what matters for a candidate. A false-alarm rate is the rate at which genuine speech is called synthetic. Compression and narrowband audio push it up sharply, and both are routine features of a video call on an ordinary connection. Codec augmentation during training closes much of the gap, which is why a well-built system can be far better than these figures — but no hiring vendor currently publishes the condition-by-condition breakdown that would let you tell.

Where that leaves processed audio

If you use anything that modifies your outbound audio, it is worth separating two risk classes, because they are not the same.

Noise suppression and echo cancellation are now built into Zoom, Teams, Meet, and most operating systems, and are near-universal among remote candidates. They remove signal rather than generating it. They are also exactly the kind of processing that pushes audio away from the clean studio recordings detectors are often trained on. The research above does not report results specifically for denoising or enhancement, so the honest position is that the effect is uncharacterised rather than known to be safe.

Accent conversion and voice conversion are a different matter. Real-time accent translation, of the kind Sanas sells to contact centres, works by resynthesising speech in a different voice. That is synthesis by construction, and a synthetic-speech detector that flags it is not malfunctioning. Candidates sometimes reach for these tools believing they are cosmetic, in the same category as a noise filter. For a system scanning the audio channel for generation artefacts, they are not. The same applies to text-to-speech used as assistive technology — which is a strong reason to raise it with the employer in advance rather than discover the flag afterwards, and to treat it as an accommodation request rather than a configuration detail.

What the law actually gives you

Voice checks sit in a more developed legal area than most interview AI, because voiceprints are named biometric identifiers in several US state statutes, including Illinois' Biometric Information Privacy Act.

The important recent development is about scope. In Delgado v. Meta Platforms, a federal court in California denied Meta's motion for summary judgment on a BIPA voiceprint claim, rejecting the argument that the statute reaches only voice data actually used to identify someone. Reporting on the ruling summarises the court's reasoning as turning on capability rather than use, and notes that the court declined to define precisely where voice data becomes a voiceprint, and rejected the idea that a formal enrolment step is required before the statute can apply (HR Daily Advisor, 28 July 2026). The practical reading for a candidate is that an employer running voice analytics over interview audio may be inside a biometric statute's scope even if nobody set out to build a voiceprint, which makes a written question about it a reasonable thing to ask.

What that does not give you is a right to avoid the check. Statutes of this kind are built around notice, written consent, retention limits, and destruction policies, not around a veto on the format of an interview. Asking what is captured and how long it is kept is well grounded. Asking to be exempted generally is not, and the same asymmetry runs through what candidates can realistically request about identity verification in remote hiring.

Practical steps before a voice interview

• Establish which layer is running. Ask whether the interview is recorded, transcribed, analysed for identity or authenticity, and whether any third-party tool is attached. A tool that joined your call as a participant is observable in the participant list; what a meeting-bot proctoring agent can and cannot see sets out the architectural limits of that arrangement.
• Give the audio the best conditions you can. A wired connection, a headset microphone close to your mouth, and a quiet room reduce the compression and noise that push detectors toward false alarms. This is the one variable genuinely under your control.
• Disclose assistive audio tooling in advance, in writing. If you rely on text-to-speech or similar, raise it before the interview and ask for it to be recorded on the file.
• Do not use accent or voice conversion on a hiring call. The benefit is contested and the artefact profile is the thing being scanned for.
• Ask the retention question early. Where a voiceprint may have been created, how long it is held and when it is destroyed are the questions the statutes are actually built around.

Where a listening assistant sits in all this

It is worth being precise about which tools this detection layer is aimed at, because candidates conflate categories that behave very differently.

Voice verification and synthetic-speech detection examine the audio you send into the call. Their target is a different person speaking for you, or a machine speaking as you. A desktop assistant that listens to the interviewer and shows you text is not in that path. Control is a desktop AI interview assistant for Windows and macOS that transcribes the live conversation in real time and renders answers in an overlay driven by global hotkeys (trycontrol.app); nothing in its documented feature set generates or alters your outbound audio. The risks attached to that category of tool are about screen capture, focus behaviour, and the delivery signals a human reviewer notices — not about voiceprints.

That is a distinction about mechanism, not an assurance. A tool that leaves the audio channel untouched is outside the scope of the checks described here; it remains inside the scope of everything else an interview observes, and of the employer's own rules about outside assistance.

The honest summary

If an interview is running voice checks, the question "is this the same person as last round" is one it can answer with reasonable confidence. The question "is this audio machine-generated" is one the published evidence says it answers much less reliably on real call audio than the product pages suggest, with false-alarm rates on genuine speech rising substantially under compression and narrowband conditions. Those two facts point the same way for a candidate: the useful work is making your audio clean and your tooling disclosed, not trying to outguess a classifier.

Before your next remote interview, the broader version of this exercise — what is captured, by what, and for how long — is worth doing once properly. Our privacy checklist for candidates using AI interview tooling walks through it end to end.