How to verify if an audio is generated by AI.

A plain-English guide to checking whether a voice clip is real or synthetic, by ear and with a detector.

To verify whether an audio clip was generated by AI, save the original file, listen for unnatural evenness in pace and breathing, then run the clip through a detector. The detector returns a probability the speech is machine-generated, a confidence level, and the likely source model, usually in under half a second. Treat a low-confidence result as inconclusive, not clean.

What verifying actually means

Verification is not a yes or no guess. It is producing a defensible answer to one question: how likely is it that this speech was produced by a synthesis model rather than a person. A good verification has three parts. A probability you can cite, a confidence level that reflects the quality of the audio, and, where possible, the name of the generator that made it. If you only have a gut feeling, you have not verified anything yet.

This matters because the human ear is no longer a reliable instrument for current voice models. In blind listening, people do little better than chance on clean synthetic speech, which is why a repeatable method beats intuition every time a real decision rides on the answer.

Field	What it means	How to read it
Probability	Likelihood the speech is machine-generated	Never rounded up; weigh with confidence
Confidence	How much the audio gave the model	Low confidence means inconclusive, not clean
Model	The recognized generator	"Unknown synthesis" when unattributable

What you can catch by ear

Listening is not proof, but it is a fast first filter and it tells you when to slow down. Play the clip twice, once for content and once purely for sound, and check for these tells.

1. Unnatural evenness. Synthetic speech often holds a steady pace and volume with none of the small stumbles, restarts, or drifting energy of a real speaker.
2. Breath and mouth sounds. Missing breaths, or breaths placed too regularly, are a common signature. Real speech has irregular inhale points and faint lip and saliva sounds.
3. Flat or looping prosody. Listen to how emotion rises and falls. Models can repeat near-identical intonation across sentences that a person would vary.
4. Room and background. A voice with no room tone, or background noise that never changes, can indicate a clip assembled rather than recorded.
5. Edges and joins. Clicks, abrupt cutoffs, or a tone that shifts mid-sentence can mean the audio was stitched from generated segments.

Any one of these is a reason to verify properly, not a conclusion on its own. Compression and bad microphones produce some of the same artifacts, so treat the checklist as a prompt to run the file, not a verdict.

When something actually matters, I treat verification as evidence handling, not a vibe check. I save the original file, I run it, and I read the whole verdict rather than the headline number, because a high probability at low confidence is not the same claim as a high probability at high confidence. If the audio is short or compressed, I try to get a cleaner or longer sample before I conclude anything. The point is not to feel certain; it is to produce an answer I could show an editor or a court and have them reach the same result.

Verify it in four steps

1. Save the original file. Download or export the actual audio. Do not screen-record it or play it into another microphone, which adds a second layer of compression and destroys the very artifacts a detector reads.
2. Run it through a detector. Open the detector and drop the file in. It reads the acoustic fingerprint and returns a result in roughly half a second.
3. Read the whole verdict. Look at the probability, the confidence level, and any model attribution together. A high probability at low confidence is not the same as a high probability at high confidence.
4. Save it if it matters. If the answer will inform a story, a filing, or a report, save the verdict to keep a permanent citation URL and a one-way fingerprint of the file you checked.

How the detector reads the audio

Real and synthesized voices leave different traces across spectral, temporal, and prosodic dimensions. The detector is trained on large numbers of examples from both sides and is retrained monthly as new synthesis systems ship, so coverage keeps pace with the market. It covers more than 24 named generators in production, including ElevenLabs, Resemble, PlayHT, and OpenAI's text to speech, and it names the source where it recognizes one. The full method is in our detection methodology, and you can see per-generator breakdowns under the detect pages.

Reading the verdict correctly

The output has a fixed shape, and reading it well is half of verification. The probability is how likely the speech is machine-generated, from 0 to 1, and it is never rounded up. The confidence is high, medium, or low, driven by audio quality and clip length. The model is the recognized generator, or "unknown synthesis" when the source cannot be attributed. There is also a methodology version stamp so a verdict can be reproduced later.

The single most common mistake is treating low confidence as a clean result. It is not. Low confidence means the audio did not give the detector enough to work with, so the honest read is inconclusive: get a longer or higher-quality sample if you can, and do not act on the clip alone. For provenance beyond detection, content-authenticity standards like C2PA aim to attach tamper-evident origin data to media, which complements a detector rather than replacing it.

Where verification is hard

Heavily compressed audio, such as phone calls and low-bitrate messaging voice notes, lowers confidence. So do very short clips, under about half a second of clear speech. In those cases the detector reports lower confidence rather than guessing, and it tells you when it cannot tell. No detector should claim certainty, and any tool that offers a guaranteed 100% verdict is overstating what the science supports. If the stakes are high and confidence is low, combine the detector with old-fashioned verification: call the person back on a number you trust, and if money or safety is involved, report it to the US Federal Trade Commission at reportfraud.ftc.gov or your national equivalent.

Frequently asked questions

Can you tell if audio is AI just by listening?

Not reliably. Listening catches obvious tells like too-even pacing or missing breaths, but people perform close to chance on clean synthetic speech. Use your ears as a first filter, then confirm with a detector before acting.

What file should I upload to verify a voice?

The original audio file, in a format like MP3, WAV, M4A, or WebM, or the audio track of a video. Avoid re-recording the sound by playing it into another device, because the extra compression removes the artifacts the detector relies on.

How long does verification take?

The detector returns a verdict in roughly half a second for a typical clip. Reading the result properly, including confidence and any model attribution, takes a few seconds more.

What does a low-confidence result mean?

That the audio was too short, noisy, or compressed for a firm answer. Treat it as inconclusive rather than clean, and try to obtain a longer or higher-quality sample before drawing a conclusion.

Is the verdict good enough to cite?

Yes. Saving a verdict produces a permanent citation URL, a methodology version stamp, and a one-way fingerprint of the file, so an editor, a court, or a board can check exactly what was analyzed.

We return a probability with a method behind it. Not a verdict on the speaker. A verdict on the audio.