Can people tell an AI voice? I tested my cloned voice on my family

Written by Inam, who built VoiceClone. I use my own trained voice for my videos, and my other engineering notes are on designesh.com. Last updated 16 September 2026.

You might hear a cloned voice online and wonder if people can actually tell it is AI. Or you might want to use a clone of your own voice for your videos and wonder the same thing. Most of the time people cannot hear the difference on a short clip with a good clone. This used to be only my opinion, but researchers have now measured it.

A 2025 study published in PLOS One showed this. Listeners called voice clones human 58% of the time, and they called real human recordings human 62% of the time. That gap is so small that they were basically guessing. So the useful thing to know is which AI voices pass, how long they pass for, and who is listening.

A well trained clone on a short clip usually passes, even with people who know your voice well. But a quick instant clone reading for ten minutes usually gets caught.

What happened when I played my cloned voice to my own family?

I tried the hardest test I could think of. With my trained clone, I generated a sentence and walked into the living room. I played it to my family and just told them to listen to a voicemail I was about to send. Nobody reacted at all, because they just thought it was me talking.

Then I told them the computer made it. They were not amazed, but they showed mild disbelief. They asked to hear it again, and they listened closely to find the join. After two listens, they still did not find it.

I did that test on my trained voice. It scores about 96 percent on the speaker match scale I use inside VoiceClone. A speaker match score is generated by a speaker recognition model. This is the type of software used for voice ID, and it compares two recordings to give a number for how close they are to being the same person. My instant clone of the same voice scores around 80 percent. Strangers accept a clone at 80, but people close to me hear that something is slightly off, just like you notice a friend wearing someone else's jacket.

What does the research say about people spotting AI voices?

You should know about two studies. Both of these studies tested normal listeners instead of audio experts, and they were done two years apart.

University College London did the first study in 2023. The researchers played real and fake speech in English and Mandarin to 529 people. The listeners caught the fake speech only 73% of the time. They also heard examples of fake speech beforehand as training, but this improved the score only slightly.

Queen Mary University of London did the second study, and it was published in PLOS One in September 2025. This is the study that really matters for voice cloning. The researchers cloned real speakers with ElevenLabs' instant cloning, and they used around 4 minutes of recordings for each voice. They asked listeners to judge each clip as human or AI. They played these clips next to real recordings, and also next to generic AI voices that copy nobody.

You can read the table like this. A generic AI voice gets caught most of the time. This happens because it does not sound like a real person, and listeners notice that. But a clone of a real voice gets called human almost as often as a real person does. So the authors concluded that clones can sound as real as human voices.

Both studies have one clear limit. The listeners only heard single sentences from voices they did not know. Nobody tested a mother listening to her own son for ten minutes. That situation is exactly where the difference between clones becomes clear.

Why is a trained clone so much harder to catch than an instant clone?

Three things decide if someone catches a clone. You can think through each of these before you publish anything.

The first thing is the type of clone. An instant clone copies your voice from a short reference clip right when it speaks. It gets your overall sound, but it misses your small speaking habits. A trained clone actually adjusts the weights of the model to match your recordings. This is why training on your own PC improved my speaker match from around 80 percent to about 96.

The second thing is how long the clip is. One sentence is easy to pass. But ten minutes without a break is hard, because a tiny mistake in one sentence will repeat across hundreds of sentences and become a pattern. Human ears are very good at hearing patterns.

The third thing is the person listening. A stranger compares your clone to nothing at all. But your mother compares it to years of hearing you speak, so she catches small things that a podcast audience never will.

What gives an AI voice away?

When people catch a clone, they almost never catch the tone. The tone is usually fine. People catch the way the voice behaves, and the things that give it away are very specific:

• The pacing is too even. Every sentence runs at the exact same speed. Real human speech speeds up, slows down, and hesitates.
• Breaths are missing, and this matters because breathing makes speech feel alive. That is why my recording guide tells you to keep breaths in your reference audio.
• A garbled word comes out smeared. This gives the clip away immediately, because no real human mouth makes that kind of sound.
• The voice says names wrongly with full confidence. A real person slows down when they are unsure of a name, but a model just says it wrong. Here is how to fix pronunciation word by word.

These are production flaws, and you can fix them. I wrote about the fixes in why AI voices sound robotic. This is also why VoiceClone makes several takes and grades every word with a speech recognizer. A speech recognizer is software that listens to the audio and writes down what it heard. This means a garbled word gets flagged on your screen before anyone else hears it.

Can an AI voice detector tell if a voice is cloned?

Detectors work sometimes, and mostly when the voice comes from the exact same company that built the detector. This surprises people, because they expect a single detector to catch every AI voice.

ElevenLabs is a good example of this, because they publish their limits openly. Their AI Speech Classifier page says the tool is built to detect speech made with ElevenLabs. It says the tool is not a general purpose detector for every AI voice provider, and it does not reliably classify audio from their own Eleven v3 model. Their newer Audio Detector looks for a watermark first. A watermark is a hidden signal written into the audio when it is generated. The tool can only find one in audio made after they introduced watermarking.

A detector that says "AI" gives you a useful hint. But a detector that says "human" proves very little. A clip made with a different tool might not have anything that the detector knows how to look for. I would never bet anything important on a single detector verdict in either direction.

Do you have to tell people when you use your own AI voice?

This is the real question hiding behind asking if people can tell. The answer depends on whose voice you are cloning.

The help page on disclosing AI content from YouTube lists "cloning one's own voice to create voice overs or dubs" as an example that creators do not have to disclose. You need the label for realistic content, like a video that makes a real person appear to say or do something they did not do.

If your own voice reads your own script, you do not need to mark it as AI on YouTube. But if someone else's voice says words they never said, you DEFINITELY HAVE TO MARK IT, and you must get their clear permission first.

My own rule outside YouTube is simple. I use my own voice or get clear permission. I also add a plain "this is AI" if a clip could pass for a real recording in a place where it matters, like a voicemail or a statement. The law is moving the same way, and I explained this in the legality article. The research says listeners cannot protect themselves by ear, so the family code word from the safety article matters a lot.

FAQ

Can people tell an AI voice from a real one?

People usually cannot tell if it is a good clone on a short clip. Listeners in a 2025 PLOS One study judged voice clones human 58% of the time and real recordings human 62% of the time. Generic AI voices and long recordings are much easier to catch.

Can AI voice detectors identify a cloned voice?

They only work partly, since detectors tend to recognize audio from their own company's engine. ElevenLabs even states that its classifier is not a general purpose detector for every AI voice provider. You should treat any single verdict as a hint instead of proof.

What gives an AI voice away most often?

An AI voice is given away by pacing that is too even, missing breath sounds, an unfixed garbled word, and names pronounced wrongly. The tone is rarely the problem. The behavior of the voice is what gives it away.

Do I have to disclose my own cloned voice on YouTube?

You do not have to disclose it for a voice over. The help page from YouTube lists cloning your own voice for voice overs or dubs as something creators can skip disclosing. But making a real person appear to say something they did not actually say definitely needs the label.

How can you test your own voice clone?

You can clone your voice in the free app. The app lets you train one voice and gives you 3 generations to check it. Use two of these on sentences you would actually say. Then play them to the person who knows your voice best, and keep a straight face.

If they catch it, the things I mentioned above will tell you exactly what to fix. If they do not catch it, you will understand why I keep my permission rule so strict.

Sources: Queen Mary University of London's "Voice clones sound realistic but not (yet) hyperrealistic" in PLOS One, UCL's "Warning: Humans cannot reliably detect speech deepfakes" in PLOS One, ElevenLabs' AI Speech Classifier page and its Audio Detector help article, and YouTube's "Disclosing use of GenAI content" help page. All read on 16 September 2026.