AI voice detector
for phone calls.
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.
of real callers are never flagged as a machine.
of AI voice calls are flagged within 10 seconds. We tested 120 voice engines, including ElevenLabs.
recordings tested: real calls, robocalls and AI voices.
How it tells a person from a machine
The detector is a small neural network that listens to the caller's audio as it comes over the phone line. We trained it on thousands of real calls, robocalls and AI voice calls, and tested it on voice engines it had never heard.
The same model runs in our voice gateway on live calls, next to transcription and content moderation. It needs only the caller's side of the call, and its score gets more certain over the first ten seconds of speech.
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 is ready half a second after the caller starts talking and arrives with their first transcribed sentence, so your system can transfer or end the call while the caller is still on the line.
Frequently asked questions
For teams adding AI voice and robocall detection to live calls.
How does the AI voice detection API work?
Stream a call to our voice gateway over WebSocket. Each transcribed sentence comes back with a robocall score from 0 to 1 and a level (low, medium or high), next to the transcript and the content moderation result. The model scores the caller’s audio, not what they say.
How fast does it flag an AI voice?
The first score is ready half a second after the caller starts talking and keeps updating for the first 10 seconds. Most AI voice calls are flagged within 3 seconds, and 94% within 10.
Which AI voices and robocalls does it detect?
AI voice agents, text-to-speech and cloned voices, from commercial services like ElevenLabs to open-source models. We tested it on 120 voice engines. Robocalls that play a recording of a real person are harder to spot from the voice alone, so the gateway also recognizes the same recording turning up across many calls.
How often does it flag a real person?
Rarely. In our tests, 99.5% of real callers were never flagged. The score starts at zero and only rises when the audio gives it a reason to.
Does it work with Twilio and other telephony providers?
Yes. The gateway speaks the Twilio Media Streams protocol, so Twilio connects directly. Other providers that can stream call audio over WebSocket can send it in the same format.
Do you need both sides of the call?
No. It scores the caller’s side on its own, so it works on inbound calls to a contact center, an IVR or a voice agent.
Can I test it before integrating?
Yes. The detector on this page runs the same model in your browser. Play the sample calls, upload your own recordings or speak into your microphone, then start free when you are ready to connect live calls.
