48,000+
Users in 84 countries
About . since January 2023 . London
The team behind the verdict.
A small team in Covent Garden building one forensic instrument for spoken authenticity. We retrain it monthly, publish the methodology, and reply to emails ourselves.
Since Jan 2023·Covent Garden, London·10 people·Retrained monthly
90,000+
Audio files analyzed
24+
Generators covered
99%
Accuracy on clean audio
Why we exist
A citable answer to "is this real?"
The open web needs an instrument, not another platform. Restrained, document-grade, honest about what it does not know.
We do not think AI voices are going to slow down. Anyone who reports, prosecutes, invests, or makes a deal based on a recording deserves a probability with a method behind it. Not a vibe. Not a label. A number you can quote.
What we are
One forensic instrument for spoken authenticity. A probability with a published method, and we keep it sharp.
What we are not
A fraud-prevention platform. A deepfake-defense startup. Enterprise SaaS. We sell one thing and do it well.
Timeline . the public record
A short history.
Only the dates we can verify. If something is missing, it is because we cannot confirm it on the record.
Jan 2023
Founded in Covent Garden, London. The detector ran its first public verdicts soon after.
Mar 2024
Methodology v2 released. Coverage extended to ElevenLabs v2 and Resemble.
2024
Adopted across newsroom, investigative, and election-integrity work.
Jan 2025
Public API opened, with a bulk processing endpoint for newsroom partners.
Apr 2026
Methodology v3.2 released. Coverage now spans 24+ synthesis systems.
How we work
Quiet. Citable. Answered.
Four operating defaults behind every verdict.
Cadence
Retrained monthly
New synthesis systems emerge constantly. We refit the detector every month and publish what changed.
Methodology
Published
The detection approach, training data sources, and limitations live on the blog. If you cite us, you can defend that citation.
Pricing
Four tiers, Free to Enterprise
Free, Starter, Business, Enterprise. No metered surprises. Cancel any time, and we refund the unused remainder.
Email
A person replies
Email reaches a real person, usually within the day. Support, press, and enterprise each have a direct address.
We started this because the alternative was watching the web get loud and stop being trustworthy. The detector is the smallest useful thing we could build in that direction.
It will not stop a determined attacker. It will not catch every novel generator the day it ships. It will give you a number with a method behind it, in under a second, and we will tell you when we are wrong.
The team
aivoicedetector.com . London
Our stance
What we will and will not say
I think the most useful thing a detector can do is be clear about its own limits, so here is where I draw the lines. We analyze voice and audio only. We do not detect manipulated images or video, beyond reading a video's audio track, and we do not monitor or intercept live calls; we work on recordings after the fact. We are not a voice-biometrics or identity-verification product, and we do not certify who is speaking, only whether the speech bears the signature of machine synthesis.
What we will do is give you a probability we never round up, a confidence level that drops honestly on short or compressed audio, the source model where we recognize it, and a permanent, citable record with a versioned methodology behind it. No black box, no guaranteed one-hundred-percent verdicts, and audio deleted within twenty-four hours unless you choose to save it. When we cannot tell, we say so. That restraint is the whole point of a forensic instrument, and it is what lets a verdict hold up when someone leans on it.