The percentage
How sure the model is that the audio was generated by a synthesis system. 97% means very likely synthetic. 50% means we cannot tell. We never round up.
The detectormethodology v3.2updated Apr 2026
Is this AI?
Drop any audio clip. Get a citable verdict in under half a second. We delete uploaded audio within 24 hours unless you save it.
Drop audio to analyze
Drag a file, paste a URL, or record straight from the page. Accepts MP3, WAV, M4A, WebM up to 25MB.
Drop file|paste URL|record 5s
Models covered24+
Accuracy99%
Median time0.48s
Retention24 hours
AccountNot required
CostFree to try
What it checks
An AI voice detector for cloned and synthetic speech.
Paste a clip and it tells you whether the voice is real or machine-generated, and which system likely made it.
This is a free AI voice detector built for one question: was this speech produced by a person, or by an AI voice generator? It reads the acoustic fingerprint of the audio, the spectral, prosodic, and timing traces that voice cloning and text-to-speech (TTS) leave behind, rather than the words being said. That is why it works across languages and accents, and why it can flag a convincing deepfake voice that sounds perfectly human to the ear.
It recognizes the major generators by name, including ElevenLabs, OpenAI, PlayHT, Resemble, Murf, and more than twenty others, so you get attribution, not just a synthetic-or-real label. Use it as a quick AI voice checker for a suspicious voicemail, a viral social clip, a podcast or interview recording, or an audio message you need to trust before acting. For a specific system, see the per-generator detection guides; to understand the signals, read how to detect AI voices and AI voice vs human voice.
How to read the verdict
A probability, not a label.
Every verdict explains what we measured, what the confidence number means, and what to do if you suspect a false positive.
01 . Confidence
02 . Attribution
The model name
When we recognize a specific generator (ElevenLabs v2, Resemble v3, PlayHT, OpenAI, and others), we name it. When we cannot, we say "unknown synthesis" and stop.
03 . What we keep
Nothing, by default
Uploaded audio is deleted within 24 hours. Verdicts you cite externally get a permanent URL with the audio fingerprint, never the audio itself.
FAQ . the detector
Questions we hear often.
Short answers. If yours is not here, the team replies to email same day.
How long does a clip need to be?
About half a second of clear speech. We accept files up to 25MB; anything longer than 5 minutes we sample down for speed.
What languages do you cover?
All major languages. Detection is acoustic, not linguistic, so accent and language affect us less than they affect transcription models.
How often are you wrong?
On clean audio (no compression, no background noise) we report 99% accuracy. On heavily compressed audio, expect lower confidence. We always show the number; we never claim certainty we do not have.
Can I use this for journalism or legal evidence?
Yes. Each verdict has a permanent URL and an APA-style citation. We publish the methodology so opposing counsel or an opposing editor can verify the test.
Do you have an API?
Yes, included with the Starter plan and above. See pricing. Enterprise customers can request a custom retention policy and dedicated retraining.
What happens to my audio?
Deleted within 24 hours. If you save a verdict to your dossier, we keep an audio fingerprint (a one-way hash), not the audio file itself.
Is there a free AI voice detector?
Yes. This one is free for a single verdict with no account or card. You see the probability, the confidence level, and the recognized model. Higher volumes and API access begin on the paid plans.
Can you detect an AI voice in a WhatsApp or phone recording?
Often, yes, though phone lines and messaging apps compress audio, which lowers confidence. The detector reports that confidence rather than guessing, so a compressed clip is not mistaken for a certain result. A clearer or longer sample improves accuracy.
How do I check if a voice is AI-generated?
Save the original audio file, drop it into the detector, and read the full verdict: the probability, the confidence level, and any named model. Avoid re-recording the sound by playing it into another device, which adds compression and removes the artifacts the detector relies on.
We keep as little as possible. The clip you upload is deleted within 24 hours. Saved verdicts retain a one-way audio fingerprint, never the audio file itself.
If you ever need us to delete your account and everything tied to it, email support@aivoicedetector.com and it is done within a day. We have never sold a single record. We have never had a breach.