Roughly 20 seconds of recorded speech is enough for current AI tools to produce a voice that listeners accept as the original speaker. The claim comes from computer scientist Hany Farid, a professor at the University of California, Berkeley, and cofounder of the security firm Get Real Security. He made it in an episode of Science News’ podcast The Deep End devoted to voice cloning.
Make AI work for you: The Neuron delivers the latest AI tools, tutorials and news straight to your inbox. Join 700,000+ readers and catch up in 3 minutes a day.
We may earn a commission if you sign up.
Farid argued that the barrier to impersonation has collapsed to a keyboard and an internet connection. Anyone with a short clip pulled from a voicemail, a video or a social post can generate a voice and a video of almost anybody saying almost anything. In his telling, the technical hurdle has moved from the attacker to the listener, who now has to decide whether a familiar voice can be trusted.
Science News host Laura Sanders presented the 20-second figure in a podcast episode on voice clones. The episode cites no dataset or methodology for the number, so it stands as an expert’s estimate rather than a measured threshold, and the research behind it points to a similar order of magnitude.
A 20-second claim from a deepfake researcher
Farid described lab listening tests in which people could not reliably separate synthetic voices from real ones, a performance he compared to flipping a coin. He also reported a personal case: someone cloned his voice to phone an attorney he was working with, and the lawyer grew suspicious after about three minutes and hung up. Even a trained skeptic was nearly fooled for a short call, which suggests how little cover a clone needs.
The same interview ran through the fraud catalog that cloned voices now feed: romance scams, fake kidnapping ransom calls, impersonated world leaders, bogus credential resets aimed at employees and North Korean IT workers disguising their identities. Farid also pointed to deepfake political ads compiled for the 2026 midterms, and the episode played clips of New York politicians Zohran Mamdani and Kathy Hochul.
Berkeley’s listening experiment and its short clips
Farid and colleagues Sarah Barrington and Emily Cooper published the underlying experiment in Scientific Reports, and a preprint abstract on arXiv sets out the main results. In the identity task, participants judged an AI-generated voice to be the same speaker as its real counterpart about 80% of the time. In the naturalness task they picked out AI voices only about 60% of the time, against a 50% chance baseline.
The published paper in the PubMed Central archive is explicit about how short the source audio was. For each of 220 speakers, a cloned voice model was built from the first two prompts using ElevenLabs’ instant cloning tool, and real clips in the test averaged 5.37 seconds. The study enlisted 604 listeners recruited through Prolific, and the authors concluded that human participants cannot consistently identify recordings of AI-generated voices.
A Berkeley news release on the study quotes Barrington saying that humans can tell something is fake only 60% of the time, and that the synthetic voices are perceptually realistic enough to fool a person. The paper does not state a 20-second threshold; it shows that a sample on the order of seconds is enough to build voices that pass.
Scam calls and the FTC’s advice
Federal regulators were warning about the mechanism before the tools became cheap. In a consumer alert, the Federal Trade Commission explained that a scammer needs only a short audio clip of a relative’s voice, which may be lifted from content posted online, plus a voice-cloning program, and the call that follows will sound like the loved one. The agency’s advice is to hang up and call back on a number already known to be the person’s, and to treat requests for wire transfers, cryptocurrency or gift-card numbers as red flags. Reports go to ReportFraud.ftc.gov. The FTC consumer alert on AI-enhanced family emergency scams dates to 2023, so the warning predates the 20-second figure by years.
Voices lent on purpose: ALS patients and griefbots
The episode also examines uses most people would welcome. Jules Rodriguez, who has ALS, communicates through a clone of his own voice, and he said hearing it made him feel he had reappeared. Performer Patrick Darling said he felt physically connected to an AI voice. A companion installment on brain implants that decode speech shows the same technology feeding assistive devices.
The ethical questions concern what happens afterward. Karina Nagin of Bridging Voice, who works with people losing speech, said she worries about deepfakes in news but sees benefits for patients, and she raised the question of who owns a voice after death. Dylan Thomas Doyle, a Northwestern University researcher who studies death and AI, described a tension between healthy continuing bonds with the dead and becoming stuck in grief, and said there is no good data on how many people use griefbots.
Sanders noted that rules vary by state, country and company, and that clear guidelines do not yet exist. What remains unresolved is the one fact the episode could not supply: a published, reproducible test that pins down the minimum number of seconds a modern cloning model needs to clear a listener’s ear.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
Make AI work for you: Get The Neuron →
We may earn a commission if you sign up.
More from Morning Overview
- Regulators cleared the first U.S. small modular reactor, 4 months early
- Lake Powell sank to a record 3,517 feet, nearing the level that stops Glen Canyon Dam’s turbines
- Seven of Earth’s nine planetary boundaries are breached, and all seven are worsening
- The NSA says three phone features should be off whenever you aren’t using them
