0

Three seconds of audio can clone a voice. How detection fights back

In June 2023, an Arizona mother named Jennifer DeStefano told a US Senate subcommittee about a phone call she had received earlier that year. She heard what sounded exactly like her 15-year-old daughter, sobbing, and then a man who said he had her and demanded $1 million. Her daughter was safe the whole time. The voice, she believed, had been cloned with AI.

How much audio does it take to fake someone's voice that convincingly? Less than most people think.

What three seconds can do

In January 2023, Microsoft researchers described a speech synthesis model called VALL-E. According to the paper's abstract, it could speak any text in the voice of an unseen speaker using just a 3-second recording as a prompt, and it preserved the emotion and acoustic environment of that recording.

It works much like a large language model. The researchers used an off-the-shelf neural audio codec to turn speech into a sequence of discrete "audio tokens," then trained a language model on 60,000 hours of English speech to predict the next audio tokens given some text and a short voice sample. The 3-second prompt tells the model the speaker's timbre, accent and habits, and the model continues in that style.

VALL-E itself was not released to the public. In the years since, open-source models and commercial services with similar abilities have spread from research labs to ordinary laptops and phone apps. And scammers have plenty of raw material: short videos, voice messages, livestream recordings. Police in Beijing warned in September 2026 that some scammers now place calls and stay silent, waiting for the victim to say "hello" or "who is this?" so they can capture a sample.

Besides text-to-speech systems like VALL-E, there is a second family of tools called voice conversion. The scammer speaks into a microphone and software turns their voice into the target's voice in real time. Text-to-speech suits a pre-recorded voice message; voice conversion lets a caller answer questions live, which makes a conversation far more convincing. The ASVspoof dataset described below includes both kinds of attack.

Can people hear the difference?

Not reliably. In a 2023 study in PLOS ONE, researchers at University College London played genuine and synthetic speech to 529 participants, in both English and Mandarin, and asked them to pick out the fakes. Listeners identified the deepfakes correctly only 73% of the time, with no difference between the two languages. Giving some participants examples of synthetic speech in advance improved their scores only slightly.

Seventy-three percent sounds decent until you picture a real scam call. The study's participants knew they were hunting for fakes. A scam victim usually isn't expecting one, and the caller creates panic ("there's been an accident," "I've been arrested," "send money now"). People under stress are less likely to notice small oddities in a voice.

That pressure is deliberate. Scam calls of this kind tend to follow the same script: something terrible has just happened, you must act right now, and you must not tell anyone else. Each part has a job. Urgency stops you thinking, and secrecy stops you checking with the one person who could expose the lie. Requests to pay by wire transfer, gift cards or cryptocurrency are another warning sign, since that money is hard to trace or recover. Once you know the pattern, the shape of the call can tip you off even when the voice sounds perfect.

How much can machines help?

Researchers have run the ASVspoof challenge series since 2015 to compare algorithms that catch spoofed speech. The 2019 edition's dataset, published in Computer Speech and Language, covers three kinds of attack: speech synthesis, voice conversion and replayed recordings, and it has become one of the field's standard benchmarks. Detectors look for traces that synthetic speech leaves in the spectrum, such as unusual energy patterns in certain frequency bands or unnatural phase behavior, details the human ear rarely notices.

The weak spot is generalization. A 2025 survey in ACM Computing Surveys reviewed more than 200 papers on speech deepfake detection published through March 2024 and gives generalizability a dedicated discussion. It is a well-known difficulty in the field: a detector trained on known synthesis methods often performs worse on a new model it has never seen. Each new generation of voice synthesis sends detection back to catch up. Some phone makers and carriers have started building scam-call warnings into their software, but these mostly flag suspicious numbers and calling patterns and do not analyze the voice itself, so they are no guarantee against a cloned voice coming from a familiar number.

The phone network makes things harder still. Ordinary calls compress audio and strip out much of the high-frequency detail, which is exactly the kind of detail detectors rely on. That is an inference from how detectors work rather than a measured result, but it suggests that spotting a cloned voice live on a phone call is much harder than analyzing a clean recording in a lab.

The scams are real

The FBI warned in a December 2024 public service announcement that criminals generate short clips of a loved one's voice to impersonate a relative in a crisis and demand immediate payment or ransom. The Federal Trade Commission had issued a similar consumer alert about AI-enhanced "family emergency" scams in 2023.

Similar cases appear in China. According to a People's Daily report republished by Xinhua in April 2025, a department manager at a company in Yinchuan received a video call from someone who looked and sounded like his boss and said money was needed urgently. Within a little over ten minutes he transferred 200,000 yuan. He discovered the fraud that afternoon when he reported to his actual boss in person.

How to protect yourself

Since listening carefully won't settle it, the reliable approach is to stop relying on the voice and check through a channel you trust.

The first recommendation in the FBI's announcement is to create a secret word or phrase with your family and ask for it whenever a call seems suspicious. It is simple, but it has to be set up in advance; once the panicked call comes, it's too late to agree on one. If you have elderly parents, agree on it in person and repeat the rule: any call asking for money, ask for the word first. It helps to rehearse once, the way families practice a fire drill, so that the response feels automatic when panic sets in. Older relatives are frequent targets partly because scammers expect them to be less familiar with AI tools, and partly because a call from a "grandchild in trouble" hits hard. Talking about this before it happens, without making anyone feel foolish, is one of the most effective protections there is.

The second is to hang up and call back. However urgent the caller sounds, end the call and dial the number you already have saved for that person, or call another relative to check. Caller ID can be spoofed, but it is much harder for a scammer to intercept a call you place yourself. The FBI gives the same advice for calls claiming to be from a bank: hang up, look up the official number, and call it directly.

The third is to give scammers less to work with. With unknown callers who stay silent, you don't have to speak first; Beijing police suggest waiting three seconds and hanging up if nobody talks. The FBI also suggests limiting public posts of your voice and making social media accounts private where you can.

Banks offer another layer. Many let customers set up delays or confirmation steps for large or first-time transfers, and some will call to check unusual payments. Turning these on for an elderly relative's account buys time, and time is exactly what a scammer is trying to take away. If money has already been sent, contact the bank immediately and report the fraud; in the US, the FBI's Internet Crime Complaint Center at ic3.gov handles these reports.

Rules are catching up too. China's deep synthesis regulations, in force since January 2023, require services that let users edit someone's face or voice to tell users to inform that person and get their separate consent. Its AI content labeling rules, in force since September 2025, require synthetic audio to carry labels such as spoken notices. Legitimate platforms follow such rules and scammers don't, so they reduce casual misuse of people's voices but won't stop an organized fraud ring.

Finally, if you are still unsure, ask something only the real person would know and that isn't online, like what you ate the last time you met. A model can copy a voice. It can't copy a memory.


Further reading: The studies cited here appeared in PLOS ONE, Computer Speech and Language and ACM Computing Surveys, among others. For public data on computer science journals, see the CrushSCI computer science journal ranking.

References

  1. Wang C, et al. Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers. arXiv, 2023. doi:10.48550/arXiv.2301.02111
  2. Mai K T, et al. Warning: Humans cannot reliably detect speech deepfakes. PLOS ONE, 2023. doi:10.1371/journal.pone.0285333
  3. Wang X, et al. ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Computer Speech & Language, 2020. doi:10.1016/j.csl.2020.101114
  4. Li M, et al. A Survey on Speech Deepfake Detection. ACM Computing Surveys, 2025. doi:10.1145/3714458
  5. Federal Bureau of Investigation. Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud. Public Service Announcement I-120324-PSA, December 3, 2024.
  6. DeStefano J. Testimony before the US Senate Judiciary Subcommittee on Human Rights and the Law, June 13, 2023.

By the CrushSCI editorial team. AI tools assisted with research and drafting; facts and citations were checked by a human editor.


All rights reserved

Viblo
Hãy đăng ký một tài khoản Viblo để nhận được nhiều bài viết thú vị hơn.
Đăng kí