Guide · 8 min read · Updated September 2026

How to tell if a voice message was cloned by AI

A cloned voice is the cheapest convincing fake available today. Thirty seconds of someone talking — a voicemail, a podcast appearance, an Instagram story — is enough material, and the tools are free. The people targeted are rarely public figures; they are parents receiving a distressed call from a child's number, finance staff told by a familiar voice to release a payment, and members of group chats forwarding a clip nobody has questioned. Audio also happens to be the hardest medium to check automatically, which makes knowing the manual tells genuinely worth your time.

By Ira Peoples

1. Listen for the breathing

Human speech is built around the need to breathe. Talk for twenty seconds and you will take breaths, and those breaths land in predictable places: at clause boundaries, before long sentences, after emphasis. They also get audibly shorter and sharper as a speaker becomes agitated.

Synthetic speech either omits breathing entirely or inserts breath sounds as decoration, placed at intervals rather than at the points where a lung would actually need air. Listen specifically for a long, fluent sentence delivered without a single intake of breath. Real speakers almost never manage this, and a panicked one never does.

2. Find the room

Every real recording contains the space it was made in. A kitchen has hard reflections off tile and glass; a car has road noise, indicator clicks and a specific boxy resonance; a street has traffic that moves relative to the microphone. That ambience is continuous, and it sits underneath the voice rather than beside it.

Cloned audio is generated dry and then, at best, has noise laid over the top. The giveaway is what happens in the gaps: real room tone continues between words at a consistent level, while added noise often ducks, loops, or cuts off cleanly at the end of the file. If the background disappears the instant the speaker stops, you are listening to two layers, not one recording.

3. Test the consonants

Plosives — the p, b, t, k sounds — are made by blocking airflow and releasing it. Close to a microphone they produce a small burst of air that registers as a thump. Sibilants, the s and sh sounds, carry a lot of high-frequency energy and vary in length depending on emphasis.

Synthetic voices tend to smooth both. Listen for p and b sounds that arrive with no physical burst behind them, and for s sounds of an oddly uniform length throughout the clip. On headphones this is easier to hear than on a phone speaker, which is worth ten seconds of your time before you act on anything.

4. Watch the rhythm, not the accent

Cloning tools reproduce timbre — the quality that makes a voice recognisable — far better than they reproduce timing. What they struggle with is the messy rhythm of unscripted speech: false starts, repeated words, a sentence abandoned halfway, the pace changing because the speaker is thinking.

A clip where every sentence is delivered at an even tempo, with well-formed grammar and no self-correction, is doing something people under stress almost never do. Emotion is a related tell: real distress distorts pitch and volume unevenly across a phrase, while a synthetic 'panicked' reading tends to apply distress as a consistent effect over the whole sentence.

5. Note the length and quality of what you were sent

Fakes are usually short, because the longer a clone runs the more chances it has to slip. They also arrive at low quality surprisingly often, and there is usually a reason attached to it — bad signal, borrowed phone, crying too hard to speak clearly. Compression conveniently destroys most of the fine detail a listener or detector would use.

Treat a stated reason for poor audio quality as information rather than an explanation. It is the audio equivalent of a document that cannot be shown to you because it is locked in a safe.

6. The check that beats every acoustic tell: call back

None of the signals above is reliable enough to bet money on, and the good news is that you do not need them to be. Voice cloning defeats recognition, not verification. Hang up and call the person back on the number you already have stored — not the number that called you, which can be spoofed. If a relative appears to be in trouble, call another family member.

Agree a spoken password with the people who might plausibly call you in an emergency: a word that would never come up naturally, known only to the two of you. It costs one conversation and defeats every clone on the market, because the attacker has your relative's voice but not your family's shared history.

7. Recognise the pressure pattern

Nearly every voice-cloning scam shares a structure: urgency, isolation, and an irreversible payment. You are told there is no time, told not to involve anyone else, and asked to move money by a method that cannot be reversed — a transfer, gift cards, or crypto.

That pattern is more diagnostic than any acoustic analysis. A genuine emergency survives a five-minute delay and a second opinion. A scam does not, which is exactly why you are being told there is no time for either.

The short version

Absent or misplaced breathing, and a background that vanishes in the gaps, are the strongest audible tells.
Cloning copies timbre well and rhythm badly: listen for unnaturally even, well-formed delivery.
Poor audio quality with an explanation attached deserves more suspicion, not less.
Call back on a stored number, or use an agreed family password. Verification beats recognition every time.
Urgency plus secrecy plus an irreversible payment is the scam signature, whatever the voice sounds like.