How to Detect an AI Voice on a Call or Recording

University College London ran a simple test in 2023. Researchers generated 50 fake speech clips in English and 50 in Mandarin, played them to 529 people and asked which were synthetic. Listeners got it right 73 percent of the time. Training them on what to listen for improved that only slightly. The authors also pointed out that their samples came from relatively old speech models.

Those listeners knew fakes were coming. On an ordinary phone call you don’t. So the useful question is not “can I hear it?” but “what can I do on a call that a cloned voice handles badly?”.

What AI voices still get wrong

Germany’s federal security agency BSI lists the artefacts its researchers see in synthetic speech (BSI):

• a metallic sound
• mispronunciation, especially of foreign words and names the model didn’t train on
• monotone delivery
• an accent or stress pattern that doesn’t match the real speaker
• unnatural noises, for example when a text-to-speech system gets an overly long sentence or a voice converter receives silence
• a noticeable delay in real-time systems, because the original speech has to happen before it can be converted

The UCL participants described their own cues. English speakers most often mentioned breathing. Mandarin speakers pointed to cadence and the pacing between words. Both groups did about equally well, which suggests there is no single magic tell.

• Tone: metallic, slightly hollow, or too clean for a phone line
• Rhythm: even pacing with no hesitation, or pauses in odd places
• Breathing: missing breaths in long sentences, or breaths that sound pasted in
• Stress: words emphasised in a way the person never would
• Names: stumbles over family names, places or foreign words
• Latency: a beat of silence before every reply, even short ones
• Background: noise that stops dead between sentences

Any one of these can have an innocent cause. Bad mobile connections make real voices sound metallic and laggy. Stress and fear change how people speak. What you are looking for is a cluster.

Tests that work during a live call

Real-time voice conversion has a weakness: it has to process speech before it can output the cloned version. Pre-generated clips have a different weakness: they can’t respond to things the scammer didn’t anticipate. Both can be tested without any equipment.

Criterion	Exposes real-time conversion	Exposes pre-made clips
Interrupt mid-sentence and change topic	partly
Ask a question only the real person can answer
Ask them to say a silly phrase you choose	partly
Ask for the family safe word
Hang up and call back on a saved number

Pre-made clips are common in emergency scams. The cloned voice cries a few sentences, then a second person takes over. If the “relative” never answers a direct question, or the handover happens the moment you ask one, that is a strong signal. Real-time systems can answer, but they tend to lag and degrade when interrupted.

The last two rows are the ones that don’t depend on audio quality at all. Our guide to a family safe word explains how to set one up, and the article on the AI voice cloning scam covers the call-back routine.

Content is often the bigger giveaway

Scammers can clone a voice. Cloning what a person knows is much harder. Many fake calls fail on content before anyone notices the sound: the “grandson” who calls you by your first name instead of Grandma, the “boss” who doesn’t know the project’s code name, the “colleague” who uses the formal greeting the real one never uses.

Pay attention to the channel as well. Would your daughter call from an unknown number instead of texting first? Would your finance director ask for a transfer over WhatsApp at 7 pm? A request that arrives through an unusual channel deserves the same suspicion as a strange-sounding voice, and it is easier to judge.

Finally, listen to what the caller wants you to do. Urgency, secrecy and an unusual payment method together form the classic scam pattern, whatever the voice sounds like.

Checking a recording or voice note

Recordings give you time, which helps. Listen with headphones, at normal speed first and then slowed down. Many audio editors, including the free Audacity, can show a spectrogram. Synthetic speech sometimes shows unusually clean gaps between words or a cut-off at high frequencies, but interpreting spectrograms reliably takes practice, and phone compression produces similar patterns. Treat it as a curiosity, not evidence.

For a more concrete signal, upload the file to the Gemini app and ask whether it was made with Google AI. Google’s SynthID check covers audio files up to an hour long (Gemini Help). Positive results mean Google’s tools produced or edited the audio. Negative results say nothing about other generators.

Some voice AI companies offer classifiers that check whether a clip came from their own service. Those are useful in the same narrow way: they can confirm their own product, not rule out everyone else’s.

Why detector tools disappoint

General-purpose audio deepfake detectors face the same problem as image and video detectors. They learn the fingerprints of the generators in their training data, and each new generation of voice models changes those fingerprints. The BSI notes that automated detection methods generalise poorly and can be fooled by deliberately adapted fakes.

Phone audio makes it worse. Calls are compressed to a narrow frequency band, which removes much of the detail detectors and humans rely on. Clips that a detector flags clearly in studio quality may look clean after passing through a mobile network.

For individuals that means: use tools as a second opinion on recordings, never as a verdict on a live call. For companies that receive many voice calls, such as banks and support centres, detection can be one layer among several. Firms building that kind of screening into a contact centre can work with a studio such as OnlineBuilders, alongside procedures that don’t depend on the model being right.

When to stop analysing and act

If the call asks for money, passwords, codes from your banking app or remote access to a computer, stop listening for artefacts. Hang up and call the person or organisation back on a number you already have. That works whether the voice is fake, real but coerced, or simply a scammer with a cold.

If you suspect a business call is a fake executive, the procedure in our guide to deepfake CEO fraud applies. For video calls, add the visual checks from how to spot a deepfake video. And if you want a family rule that settles every doubtful call in ten seconds, set up the safe word today.

Questions people ask

Is there an app that detects AI voices during a phone call?

A few security products and some phone makers offer call screening that flags suspected synthetic speech, but none is reliable enough to bet money on. Treat a warning as a reason to verify, and the absence of a warning as no information.

Can AI voices fake breathing and emotion?

Increasingly, yes. Newer models add breaths, laughter and sobbing. That is why sound-based tells are getting weaker and content-based checks, such as questions only the real person can answer, matter more.

How can I check a voice message someone sent me?

Upload it to the Gemini app and ask whether it was made with Google AI, which reads the SynthID watermark. Then check the content: would this person say this, in this way, through this channel? If it asks for money, call them.

Do voice clones work in other languages?

Yes, and that can be a tell. Germany's BSI notes that models trained on one language often mispronounce foreign words or names. If a bilingual relative suddenly stumbles over their own surname, be suspicious.