How to tell if a voice recording is AI-generated

Can you know for sure that a voice is AI-generated?

No. Neither the human ear nor software can be sure in every case. Voice synthesis and cloning tools imitate a person's tone and delivery from recordings of them, and the output improves with each generation.

That changes the useful question. Instead of "does it sound real?", ask "can I confirm some other way that this person said this?" Careful listening is good for raising suspicion. For making a decision, context is what counts.

What are the signs of a cloned voice when you listen?

Some traits show up often in synthetic audio. None is conclusive, but several together are a reason to be wary:

• Flat or oddly even delivery. The voice keeps the same pace and energy, without the rises and dips of someone who is truly upset or rushed.
• Missing breaths. Long sentences with no intake of air, or breaths that sound pasted in and always identical.
• Unnatural pauses. Silences in the middle of a phrase, or no hesitation at all where a person would naturally falter.
• Mangled names and numbers. Surnames, nicknames, acronyms, amounts or addresses pronounced strangely, or in an accent the person does not have.
• A background that is too clean. No room noise at all, or noise that switches on and off abruptly.
• Abrupt cuts. Words that start or stop sharply, or a jump in tone between one sentence and the next.

These cues have two serious limits. A well-made clone does not show them, and a phone call or a compressed voice message erases exactly the fine detail that would let you notice them. Audio with no warning signs is not verified audio.

It cuts the other way too. A real person on a bad connection, with a cold, or reading from a script can sound flat and choppy. Going by ear alone produces mistakes in both directions.

How do you verify that the caller is really who they claim to be?

With checks that do not depend on the sound. They are simple, and they work even against a perfect clone, because an impersonator does not control the other channels or what the real person knows.

The most consistent pattern in voice clone scams is not in the audio but in the request: an emergency, a great hurry, an instruction to tell nobody, and a bank transfer, gift cards, a verification code or a password. When you see that combination, verify every time, even if the voice seems unmistakable.

What do AI voice detectors do, and how far can you trust them?

They examine the recording for traits that tend to separate synthetic speech from human speech and return an estimate. They are a second opinion, not a verdict.

It pays to know their limits before relying on one:

• Noise. Music, wind or several people talking mask the signals.
• Compression. Audio that was forwarded several times, or recorded off another phone's speaker, has lost information.
• Length. A clip of a few seconds offers very little material to analyze.
• Errors both ways. A detector can flag a real voice as artificial and let a clone through.
• Lag. Synthesis tools change faster than the methods used to recognize them.

The AI audio detector answers in three levels: likely AI-generated, likely a person, or inconclusive. It also states its confidence (high, medium or low), the signals it observed and the limits of the analysis. It gives no percentages and no categorical claims, because that would promise a precision nobody has. The file is uploaded for analysis and deleted when the job finishes.

A "likely a person" result is not clearance to send money, and "likely AI-generated" does not prove anyone committed fraud. Either way, you still need to verify through another channel. The same reasoning applies to pictures, as covered in how to tell if an image is AI-generated.

What should you do if you get a suspicious voice message or call?

1. Do not act on the spot. Send no money, give out no codes or passwords, share no details.
2. End the call with any excuse and reach the person on a number or channel you already knew.
3. If you cannot reach them, call someone close to them before doing anything else.
4. If it turns out to be a scam, warn the person who was impersonated and the people around you. Others are likely to be targeted next.
5. If you already sent money or handed over information, contact your bank immediately and report it to the police or the relevant fraud authority in your country.

How do you preserve the audio as evidence?

Keep the original file exactly as it arrived. Audio re-recorded with another phone or forwarded several times loses quality and data, which makes any later analysis harder.

• Do not delete the message or the conversation it arrived in.
• Save or export the audio file from the app instead of recording the screen or the speaker.
• Write down the originating number or account, the date and the time, and take screenshots of the surrounding context.
• If it was a live call, note the time, the number and what was said right away. Rules on recording calls differ from country to country.
• Do not trim or convert the file. If you need an edited version, work on a copy.

To record that the file existed on a given date and has not changed since, you can lock it with Certify a file, or generate an Authenticity report, which brings together the hash, the timestamp and the detector's estimate. That establishes the integrity and date of the file. It does not prove who created it, does not guarantee legal value and does not replace an expert. In a serious case, voice analysis is a job for a forensic specialist.

Checklist and common mistakes

• Are they asking for money, codes or passwords, urgently or in secret?
• Did the contact come from a number or account you did not know?
• Were you able to speak to the person on a channel you already had?
• Did they answer something only they would know, or give the agreed safe word?
• Did you keep the original file and the details of where it came from?

The most common mistakes:

• Trusting it because "it was their voice." Tone no longer identifies anyone.
• Calling back the same number that called you, or one you were given during the call.
• Treating a detector result as proof, in either direction.
• Dropping your suspicion because the audio has no audible flaws.
• Forwarding the audio to many contacts for opinions and losing track of the original file.
• Picking a safe word that can be worked out from social media, such as a pet's name.