AI voice detector
for phone calls.
Pick a sample call, upload a recording or speak into your microphone. The score updates while the audio plays. AI voices and robocalls usually cross the flag line within three seconds.
Scoring runs in your browser. Nothing you upload or say leaves your device.
Choose a call below, record yourself, or drop your own recording anywhere on the page.
Warranty scam
An extended-warranty pitch read by a commercial AI voice model of the kind used for live voice agents. We wrote the script for this page, and the clip was not used in training.
genuine phone callers reach the flagged level, averaged over six sets of real calls and voicemails. No set went above 1 in 100.
of robocalls and AI voices are flagged within 10 seconds, scored by a model that never heard that voice engine. Most flags land in the first 3 seconds.
recordings tested: 1,786 real calls and voicemails, 399 FTC robocalls and 248 AI-voice calls from 7 engines.
Offline evaluation, October 2026, on 8 kHz phone audio. None of the four demo calls were used in training: the robocall comes from a campaign the model never heard, and the AI calls were generated for this page. Of 22 robocalls from unheard campaigns, 18 reached the flagged level.
How it tells a person from a machine
Synthetic voices improve every few months, so the detector doesn't try to judge how human a voice sounds. It scores the call audio itself, with a model trained on thousands of real phone calls, robocalls and calls made with AI voices.
The same model runs in our voice gateway on live calls, alongside transcription and content moderation. It only needs the caller's side of the call. Each transcribed sentence comes with a score, and the score gets steadier the longer the caller talks.
Demo audio: the robocall is from the FTC’s Project Point of No Entry recordings (public domain), via the NC State robocall audio dataset. The bank caller is from the Harper Valley Bank corpus by Gridspace and Stanford (CC BY 4.0). The AI calls were generated for this page with a frontier commercial voice model and an open-source one. Scam scripts are fictional.
Robocall and deepfake detection for live calls
Stream a call to our voice gateway over WebSocket, from Twilio Media Streams or your own telephony. Every transcribed sentence comes back with its content moderation result and a robocall score.
The first score arrives within a second of audio, so your system can transfer or end the call while the caller is still talking.
Frequently asked questions
Common questions about detecting AI voices and robocalls on phone calls.
How does the AI voice detector work?
A model trained on thousands of real phone calls, robocalls and AI-voice calls scores the caller’s audio from 0 to 1. It doesn’t depend on how natural the voice sounds or on what the caller says.
Can it detect ElevenLabs and other AI voices?
Yes. We tested it on calls voiced by 7 AI voice engines, both commercial and open-source, scoring each engine with a model that had never heard it. 74% of robocalls and AI voices were flagged within 10 seconds, most of them in the first 3 seconds.
Is the AI voice detector free? Is my audio uploaded?
It is free and runs entirely in your browser. Recordings you add and audio from your microphone are scored on your device and never uploaded.
Does it detect robocalls and prerecorded messages?
Yes. Prerecorded messages and text-to-speech are scored the same way as AI voices. Of 22 robocalls from campaigns the model never heard, 18 reached the flagged level.
How often does it flag a real person?
About 1 in 200 genuine phone callers reach the flagged level, averaged over six sets of real calls and voicemails. No set went above 1 in 100. The score starts at zero and rises only when the audio gives it a reason to.
Can I use it on live calls through an API?
Yes. The same model runs in our voice gateway. Stream a call over WebSocket (it works with Twilio Media Streams) and every transcribed sentence comes back with a robocall score next to the transcript and content moderation. It needs only the caller’s side of the call.
What audio works best?
One speaker’s side of a phone call, at least a few seconds long. A recording with both callers mixed together scores less reliably.
