Twelve readers, and twelve different ways to be wrong
Every page below answers four questions for one kind of reader: what your exposure actually is, what an analysis can add to it, what it cannot do at all, and what you owe the person on the other side of a result. The third answer is the longest one, every time.
How to pick the page that fits you
Job title is the wrong index. Two people with the same title need opposite pages, and the thing that separates them is the decision sitting in front of them. Three questions sort it quickly.
Is the audio live, or already a file? Nothing here helps while a caller is still speaking. If the requirement is to stop a call in progress, the pages for contact centres and banks explain why we are not that vendor. If the audio is a recording in a case file, an inbox or an evidence folder, you are in the right section.
Does a person get hurt if the result is wrong? If a verdict could cost someone a job, a claim, a place on a course or a reputation, read the page written from that person’s side of the table — HR, insurers, education, recruiters. Those four pages are mostly about the false positive rate, because that is the number that decides how much damage the process does when it misfires.
Are you using it, buying it, or citing it? Using is the audience pages. Buying is enterprises, which answers procurement in the order procurement asks. Citing is researchers, which is about method, versioning and whether a result can be reproduced by someone who does not trust us.
People, and the institutions that act on them
Families and individuals
A safe phrase, a callback rule and a quieter online voice — the plainest version of the advice, and the one page where the detector matters least.
Schools and universities
Oral submissions, vivas and admissions interviews: the institution writes the policy and the student carries the entire cost of it being wrong.
HR teams
The false positive rate is the number that should govern the process, because an accused employee has no evidence available that answers the accusation.
Recruiters
Interview proxying and real-time voice conversion, handled with disclosure to the candidate and no automated rejection on the strength of a score.
The newsroom, the court and the lab
Journalists
What belongs in the piece — tool, engine version, the error rate for that audio condition, the reference code — and how to handle a source who denies a recording that is real.
Lawyers
A screening document, not an expert opinion. Admissibility belongs to the court and chain of custody belongs to your client.
Researchers
Reproducibility rather than verdicts: what we publish, what is in the evaluation set, and the fact that no independent party has audited any of it.
Business and risk
Enterprise buyers
The procurement answers in writing, including the two that lose deals: no SOC 2 or ISO 27001, and quantitative accuracy results still forthcoming.
Security and incident response
Detection is post-incident forensics rather than prevention — where it belongs in a runbook, and how to preserve audio without quietly destroying it.
Banks and fintech
Not a live-call screening vendor. The after-the-fact tool for disputes, post-incident review and testing your own controls.
Contact centres
Caller verification on recorded calls, and the reverse exposure of pointing the same tool at your own agents on 8 kHz audio.
Insurers
Two errors are available on a recorded statement, and the one that harms a genuine claimant is the one your threshold should be set against.
What every page here refuses to say
None of them claim a verdict is proof. The two results are not symmetric and each page says so in its own terms: likely synthetic is the stronger finding because it needs positive evidence of generation, while likely human can equally mean that a codec, a re-encode or a forwarded copy scrubbed the evidence somewhere between the microphone and you. On the compressed material most organisations actually hold, a human verdict is closer to an absence of finding than to a clearance.
None of them quote a detection rate without the false positive rate beside it. Both are published together, on a described set, and re-measured whenever the engine changes. Quantitative results are still forthcoming, and a policy template exists for the reasoning in the meantime.
Who should not buy this
If you need a decision while a caller is still on the line, this is the wrong product — we analyse recordings after the event, and an inline screening vendor is the right answer for a contact centre. If you need to know which person is speaking, that is speaker comparison and we do not do it. If you need proof rather than a probability, no detector on the market can give you that, and one that implies otherwise is selling you a liability.
If your procurement process requires a current SOC 2 report or an ISO 27001 certificate to open a file, stop here: we hold neither, and the enterprise page says so at the top rather than at the bottom of a questionnaire.
And if the plan is to act against an individual — dismiss, decline, reject, expel, accuse — on the strength of a score, please do not buy this from anyone. The rate at which any detector is wrong about a genuine recording is not zero, and the person on the wrong side of that error has nothing they can produce that settles it. Every page here says so, because it is the failure mode that does the real damage.
The rest is straightforward. A verdict is one input. A named human decides, and can be asked to explain the decision without saying that the model told them so.
These pages describe duties that commonly arise when you analyse someone’s voice. They are not legal advice, the rules differ between jurisdictions, and you should take your own advice in the markets you operate in.
Questions about the section
Which page should I read if two of them describe me?
Read the one written from the point of view of whoever is harmed by a wrong result, then the one matching your role. A fraud investigator at an insurer gets more from the insurers page than the security page, because the claimant is the person carrying the risk. The role page usually tells you how; the harm page tells you where the threshold should sit.
Do the pages contradict each other?
They differ in emphasis by design. A newsroom is told to chase the original file because provenance beats any score; an HR team is told to weight the false positive rate above everything else. Both follow from the same limits applied to different stakes. If you find an actual contradiction, tell us and we will fix the page rather than the argument: press@aivoicedetctor.com
Is there a page for my sector that is not listed?
Not yet. Twelve is what we have written honestly; the rest would be find-and-replace, which is how audience pages usually get made and why most of them are worthless. If your use case genuinely differs, describe it through contact and it may become the thirteenth.
Before any of these pages, read the limits
The fastest way to work out whether this belongs in your process is to read what it gets wrong first and the sales copy second.
Reviewed