The standard for spoken authenticity

AI Voice Detector. Detect AI-generated voices.

Drop in any clip. Get a citable verdict in under half a second. We publish the method and keep nothing.

Open detector Chrome extension

0.48s

Median verdict

99%

Clean audio

24+

Models covered

48k+

Global users

Synthetic detected

Verdict

Likely fabricated.

Splice 0:14

Confidence

97.3%

Source

WhatsApp . 0:14

Try the detector on your audio →

Who uses it

Built for anyone making a decision on a recording.

From newsrooms breaking a story to compliance teams reviewing a flagged call. Same instrument. Same methodology.

Fig 1.1

Newsrooms

Journalists

Verify suspicious audio before publication. Cite verdicts in print. The methodology holds up in editorial review.

Cited in 12+ outlets

Fig 1.2

Federal . legal

Investigators

Chain-of-custody friendly workflow. Permanent verdict URLs. Defensible methodology in deposition.

Used in 30+ cases

Fig 1.3

Compliance

Banks & call centers

Flag CEO-impersonation fraud before the wire goes out. API plus audit log. SOC 2 reports on request.

8 enterprise customers

Fig 1.4

Individuals

Anyone curious

Got a voicemail that felt off? Drop it in. Get a verdict in half a second. No account required to try.

48,000+ users . 84 countries

What it does

Six things the detector actually does.

Same model across every surface. The differences are in how you reach it and what you can do with the verdict.

Fig 0.1

Drop & detect

Drag any audio file, paste a URL, or record straight from the browser. Verdict in 0.48s.

Fig 0.2

Model attribution

When we recognize the generator (ElevenLabs, Resemble, PlayHT, OpenAI), we name it. When we cannot, we say so.

Fig 0.3

Citable verdict

Every verdict gets a permanent URL and an APA-style citation. Quote it. Subpoena it. File it.

Fig 0.4

Dossier history

Save verdicts to a private dossier. Audit log. Shared dossiers for teams. Export to CSV or PDF.

Fig 0.5

REST API

One endpoint. JSON in, JSON out. Same model as the public detector. Webhooks for async jobs.

Fig 0.6

Chrome extension

Right-click audio on any tab and read the verdict in a side panel. Citations are one click.

By the numbers . since Jan 2023

The instrument newsrooms cite by name.

Quietly, since January 2023. No marketing budget. Just the model, kept current, and the team behind it answering email.

Users . 84 countries

48,000+

Protected from voice-based fraud and impersonation.

Files . cumulative

90,000+

Audio clips analyzed and filed since launch.

On staff . London

Researchers and engineers in one office.

Enterprise . production

Newsrooms, banks, and call centers on the API.

Three steps . half a second

From clip to citation.

No account required to run a single verdict. Sign in only when you want to save it to your dossier or cite it externally.

01 . Submit

Drop audio.

Paste a URL, drag a file, or record straight from the page. Accepts MP3, WAV, M4A, WebM up to 25MB.

25 MB max . 24 formats

02 . Analyze

The model listens.

Roughly half a second of compute. We run the clip against 24+ generator signatures and return a probability, not a label.

0.48s median . 24+ models

03 . Cite

Get the notice.

A permanent, citable verdict with confidence, attribution, and a methodology link. Quote it. Subpoena it. File it.

Permanent URL . APA citation

Coverage . retrained monthly

Twenty-four generators, named.

Every major voice synthesis system in production. New entrants added the month they ship. Methodology version stamped on every verdict.

ElevenLabs v2Voice cloning

ElevenLabs v3Voice cloning

Resemble v3Voice cloning

PlayHTTTS . cloning

OpenAI TTSTTS

OpenAI TTS-HDTTS

Murf AITTS

WellSaidTTS

SpeechifyTTS

Azure NeuralMicrosoft

Google WaveNetTTS

Amazon PollyTTS . neural

Coqui XTTSOpen source

BarkOpen source

Tortoise TTSOpen source

VALL-EMicrosoft

VALL-E XMicrosoft

NaturalSpeechMicrosoft

SunoVoice . music

Descript OverdubVoice cloning

iSpeechTTS

LovoVoice cloning

ReplicaVoice cloning

+ methodology →

Comparison . why us

What you get vs. listening twice.

The human ear is unreliable on current-generation synthesis. The detector returns a citable probability with a published method behind it.

Capability	Listening twice	Generic AI detector	aivoicedetector.com
Verdict speed	Minutes	2-5 seconds	0.48s median
Accuracy on clean audio	~60%	85-92%	99%
Model attribution	No	Sometimes	24+ generators named
Citable output	No	No	Permanent URL + APA
Methodology published	N/A	Rare	Versioned monthly
Audio retention	N/A	30-90 days	24 hours, fingerprint after
Free tier	Free	Sometimes	50 a month, no card

Developer API

Put the verdict in your stack.

The same model behind one REST endpoint. POST an audio clip, get a citable probability with model attribution in under half a second. Webhooks for bulk, JSON in and out.

Explore the API →

POST /v1/detect

curl -X POST .../v1/detect \
  -F "audio=@call.mp3"

→ 200 OK
{
  "verdict": "likely_synthetic",
  "probability": 0.973,
  "model": "ElevenLabs v2"
}

We don't think AI voices are going to slow down. We think the open web needs a single, citable instrument for asking "is this real?" and getting an answer with receipts.

We retrain monthly. We publish our methodology. Every verdict is permanent and quotable. No black box, no marketing budget. Just the model, kept current, and the team behind it answering email.

Abdellah Azzouzi

Founder . London

Read the methodology

From the blog . latest

What we found recently.

All articles →

Methodology

How to detect AI voices. Voice cloning and audio deepfakes.

A step-by-step methodology for verifying suspicious audio. Includes a fast checklist and tools.

Apr 202514 min

Casebook

What are the dangers of AI voices?

Five categories of harm. Real cases, real losses, and what actually works to prevent them.

Feb 20259 min

Casebook

Deepfakes in elections. Impact and how to combat them.

2024 was the first election cycle where AI audio reached voters at scale. What we saw.

Dec 20247 min

FAQ . common questions

Questions we hear often.

If yours is not here, email the founder. Reply within the day, every day.

How accurate is the detector?

99% on clean audio across 24+ generator models. We always report the confidence number; we never claim certainty we do not have. On heavily compressed audio (phone calls, low-bitrate WhatsApp), expect lower confidence. We tell you when we cannot tell.

How long does a clip need to be?

About half a second of clear speech. Files up to 25MB on Free and Starter, longer on Business and Enterprise. Anything past the clip limit is sampled down for speed.

What happens to my audio?

Deleted within 24 hours by default. If you save a verdict to your dossier, we keep a one-way audio fingerprint (a 256-bit hash), never the audio file itself. We never use uploaded audio to train the model unless you explicitly opt in.

Can I cite this in journalism or legal evidence?

Yes. Every verdict has a permanent URL and an APA-style citation. The methodology is published and versioned. Opposing counsel or an opposing editor can verify the test against the same methodology snapshot used at the time of your verdict.

Is there a free tier?

Yes. 50 detections per month, no card required. Voice and video clips up to 2 minutes. AI/Human verdict + TrustScore on every detection. See pricing for paid tiers.

Do you have an API?

Yes. Included with the Starter plan and above. REST endpoint, JSON in, JSON out. Same model as the public detector. Webhooks supported. Enterprise customers can request a custom retention policy and dedicated retraining cadence.