
// upload, then detect POST /api/v1/detect/image { "data": { "verdict": "likely_ai", "verdict_label": "Likely AI", "ai_score": 0.71, "source_trace": { "tool": "Midjourney" } } }
ResearchIs this voice AI? Upload a voice note, voicemail or call recording and find out whether the speech is a real person or a voice clone from tools like ElevenLabs, PlayHT, Resemble or OpenAI.
Upload a recording, and the detector scores the speech for signs of AI generation, slice by slice, then gives you one clear verdict.
Drop in an MP3, WAV, OGG or M4A file: a voicemail, a voice note, a call recording or a clip from a video. Free accounts check clips up to 2 minutes; paid plans go up to 10 minutes.
Clips longer than 6 seconds are split into segments, and each segment gets its own AI score. A real voice with a cloned sentence spliced in shows up as a run of flagged segments.
You get an overall AI score from 0 to 100, a verdict in four bands, and a segment timeline that shows exactly where the speech looks synthetic.
No meaningful signs of synthetic speech.
Mostly natural speech, with some segments worth a second listen.
More synthetic signals than natural ones. Check the flagged segments and verify the caller another way.
Strong signs of a cloned or text-to-speech voice across the recording.
One score for a whole file can hide a problem. A clip that is mostly real with one cloned sentence gets averaged away.

The detector scores every 6-second segment of the recording on its own, then reports the average and the highest segment score alongside the overall verdict.
That matters for the scams that cause real damage. A fraudster can splice a few cloned words, "it's me, send the money", into a real recording. A segment timeline points straight at the part that changed, so you know which sentence to question.
Use these alongside the detector. No single sign is proof, but two or three together are worth acting on.
Real speakers rise and fall without trying. A cloned voice often stays level, or sounds panicked in a way that never changes.
People breathe between phrases. Synthetic speech often runs sentences together, or adds breaths at oddly regular intervals.
Every word lands at the same speed, with no hesitations, restarts or filler words like "um" and "uh".
A "phone call" from a street or a car should have noise. Dead silence behind the voice is a common tell.
Names, places and numbers come out slightly wrong, or with stress on the wrong syllable.
"Don't tell anyone", "send it now", "I can't talk long". Pressure is the scam, whatever the voice sounds like.
The full guide is in how to detect AI voices. For audio deepfakes in videos and calls, read how to detect deepfake audio. Comparing tools? See the best AI voice detectors.
A few seconds of someone's voice from social media is enough to clone it. These are documented cases.

Hong Kong, $25.6 million. A finance worker sent $25.6 million after a video call in which the CFO and colleagues were deepfakes, cloned voices included.
Hyderabad, Rs 1.4 lakh. A woman lost Rs 1.4 lakh to a scammer who used AI to mimic her nephew's voice asking for urgent help.
"Mom, these bad men have me." A mother received a fake kidnapping call using a clone of her daughter's voice, with a ransom demand behind it.
Each one followed the same pattern: a familiar voice, a sudden emergency and a demand for money or secrecy. For more cases and how they work, read our guide to voice cloning scams and how AI voice cloning works.
Synthetic speech from current text-to-speech and voice-cloning systems, including these. Developers can run the same checks through the AI voice detector API.
Yes. A free account checks voice, images and video, with no card. Paid plans add volume, longer clips and the API.
An AI voice detector analyzes an audio recording and estimates whether the speech is a real human voice or AI-generated, such as a voice clone or text-to-speech. DeepfakeDetector.ai scores the recording in 6-second segments and gives an AI score from 0 to 100 with a verdict: Human, Likely Human, Likely AI or AI Generated.
Listen for flat emotion, missing breaths, unnaturally even pacing, mispronounced names and a background that is too clean. No single sign is proof, so stack a few cues, check who sent the message and why, and then confirm with the detector, which scores every 6 seconds of the recording on its own.
Yes. A free account includes 50 checks a month across voice, image and video, with no card required. Free accounts check voice clips up to 2 minutes. Paid plans start at $49 a month and add more checks, clips up to 10 minutes and API access.
Yes. Save the voice note, voicemail or call recording as an audio file and upload it. MP3, WAV, OGG and M4A are supported, which covers most WhatsApp voice notes and voicemail exports. Each 6-second segment gets its own score, so one cloned sentence inside a real recording still stands out.
Every result comes with an AI score rather than a bare yes or no, because detection is probabilistic. Results are strongest on clear recordings; heavy compression, music and background noise make any check harder, so weigh the verdict alongside who sent the recording, what it asks you to do and whether you can confirm it another way.
Yes. It checks for synthetic speech from voice cloning and text-to-speech tools such as ElevenLabs, Resemble AI, PlayHT and OpenAI voices. The result tells you whether the speech looks AI-generated and where in the recording the synthetic segments are, with an AI score from 0 to 100 and one of four verdicts.
No. The detector tells you whether speech looks AI-generated and where, but it does not name the tool or the voice that made it. Some vendors check only their own output, for example the ElevenLabs speech classifier for ElevenLabs audio, so use a vendor checker when you already suspect a specific tool.
Not in real time. The detector checks recordings. If a call feels wrong, hang up and call the person back on a number you already know, agree a family code word for emergencies, then save the voicemail or recording where lawful and upload it for a verdict.
Upload a voice note, voicemail or call recording and get a verdict, an AI score and a segment timeline.