SI detection

How to tell if a voice is SI: what ears and voice detectors can catch

A cream retro telephone handset on a dark floor with a cord, beside a cream sound wave whose right end breaks apart into drifting amber pixels
Illustration created with Claude (Anthropic).

Listeners spot a cloned SI (AI) voice only about 60% of the time, and voice detectors lose accuracy outside the lab. What studies, the FTC and the FBI say, and the checks that actually work.

How do you tell if a voice is AI? In SI (AI) terms, the honest answer in 2026 is that your ears alone will not tell you reliably, and many AI voice detectors were never built to answer the question you are actually asking. A cloned voice can be made from a few seconds of someone's speech, it can sound like your daughter, your boss or a politician, and it reaches you through a phone line that strips away much of the detail a detector would need. This guide covers what research says about human listeners, what voice detectors can and cannot do, and the checks that work when a voice asks you for money or a password.

The short version. In a 2025 study, listeners matched a cloned voice to the real person about 80% of the time and spotted it as SI only about 60% of the time. Machine detectors do better in the lab, but lose accuracy on real recordings, compressed audio and voices from generators they have not seen. Public classifiers often only recognise their own company's voices. When a voice wants something urgent, verify through a second channel instead of trying to judge the sound.

Can people tell a cloned voice from a real one?

Not consistently. The clearest recent evidence comes from Sarah Barrington, Emily Cooper and Hany Farid at the University of California, Berkeley, published in Scientific Reports on 31 March 2025. They cloned the voices of 220 speakers with ElevenLabs' Instant Voice Cloning, feeding the tool only the first two short prompts each speaker had recorded, and played real and cloned clips to 604 participants recruited online.

The results are uncomfortable. When people heard a real clip and a clone of the same speaker, they judged the two to be the same person about 80% of the time. When they were asked directly whether a single clip was real or SI-generated, they were right 67.4% of the time on real voices and 60.8% of the time on cloned ones, according to the paper's preprint. So a cloned voice passed as real roughly four times in ten, in a calm test where people knew fakes were in the mix. The authors' conclusion is blunt: relying on human perception to detect SI (AI) voice clones "is no longer consistently reliable".

The same paper sums up earlier work: older studies with fewer speakers, shorter phrases and less advanced synthesis found people right about 70 to 80% of the time. The trend is not in the listener's favour. Our guide to SI scams with cloned voices, fake photos and fake profiles covers one of those earlier studies, from University College London, and what it means on a real phone call.

Why is a cloned voice so hard to hear?

Several things work against you at once.

  • Very little audio is needed. The Berkeley clones were built from two short recordings per person. Anyone who has posted a video, a voicemail greeting or a podcast clip has given away that much.
  • Identity is what you listen for. On the phone you ask yourself "is this my brother?", not "is this a synthetic waveform?". The study found that identity is exactly where clones are most convincing.
  • Phone audio hides detail. A call, a voice note in a messaging app or a clip re-shared on social media is compressed. The small irregularities that might give a clone away are the first thing compression removes.
  • The script is built to rush you. Distress, secrecy and a deadline leave no room for careful listening.

The FBI still advises people to "listen closely to the tone and word choice" of a caller, in its public service announcement on generative SI fraud of 3 December 2024. That is a sensible first alarm: a relative who suddenly uses words they never use deserves suspicion. But a clone that sounds right proves nothing.

What does an AI voice detector actually do?

An SI voice detector, or AI voice detector, is a classifier: a model trained on examples of real and synthetic speech that outputs a probability that a clip was generated. Researchers call this spoofing or audio deepfake detection. Tools fall into three groups.

Classifiers that only know one generator

Some SI companies publish a checker for their own output. ElevenLabs, for example, offers a free AI speech classifier that "estimates the probability that an audio sample was generated with ElevenLabs AI voice technology". Its own page is clear about the limits: it "cannot detect audio from other providers", it "does not reliably classify audio generated with the Eleven v3 model", and it analyses only the first minute of a sample. Useful if you suspect that one service. A low score tells you nothing about the dozens of other voice generators.

General detectors

Other tools try to recognise synthetic speech from any source. They face the problem described below: accuracy measured on one set of voices does not carry over neatly to the next generator.

Real-time call screening

A third group works on live calls, mostly for banks and call centres. The US Federal Trade Commission gave a recognition award in its Voice Cloning Challenge to Pindrop, whose technology evaluates incoming calls in two-second chunks and flags possible deepfakes as the call happens. These systems sit inside a company's phone infrastructure, not on your own phone.

How accurate are voice detectors?

In a controlled benchmark, the best can look excellent. Outside it, the picture changes. Two pieces of research show the pattern clearly.

ASVspoof is the long-running academic challenge for spoofed and deepfake speech. Its fifth edition, described in an overview paper accepted by the journal IEEE TASLP (preprint first posted in January 2026), used a new crowdsourced database with many more speakers, varied recording conditions and a mix of recent and older speech generators. Fifty-three teams submitted systems. The organisers write that "many solutions perform well", but that performance degrades under adversarial attacks and when the audio goes through neural encoding or compression schemes, which is exactly what happens to speech sent over modern apps.

The generalisation problem had already been measured in 2022. Nicolas Müller and colleagues collected 37.9 hours of real-world recordings of celebrities and politicians, 17.2 hours of them deepfakes, and tested detectors from earlier research on this "in the wild" audio. Their Interspeech 2022 paper reports that these detectors performed poorly on real-world data, with performance degradation of up to one thousand percent, and suggests the field may have tailored its solutions too closely to the ASVspoof benchmark. Deepfakes, they concluded, may be "much harder to detect outside the lab than previously thought".

The lesson is the same one we found for images and text in our review of whether SI detectors are accurate: an accuracy figure belongs to the test set it was measured on. A new generator, a phone codec or a clip recorded off a speaker can all push a detector outside what it learned. Treat a voice detector's score as one clue, never as a verdict.

Put side by side, the options look like this:

  • Your own ears: a first alarm if the words or tone feel wrong. In the Berkeley study, clones passed as real about four times in ten.
  • A single-company classifier: whether a clip probably came from that one company's generator, and nothing about the others.
  • A general voice detector: a probability, weaker on compressed audio, new generators and real-world recordings.
  • Calling back on a number you already know: whether the real person or organisation actually made the request. It is the only check here that answers that question directly.

What did the FTC Voice Cloning Challenge find?

In 2024 the FTC ran a public challenge for ideas to protect consumers from voice cloning, and announced the winners on 8 April 2024. Three entries shared $35,000, and they illustrate three very different strategies:

  • AI Detect, from OmniSpeech, uses SI algorithms to "distinguish the subtle differences between genuine and synthetic voice patterns". This is detection after the fact.
  • DeFake, from Ning Zhang at Washington University in St. Louis, adds "carefully crafted distortions" to voice recordings that are imperceptible to the human ear, to make accurate cloning harder in the first place. This is protection before the fact.
  • OriginStory, from a team including Visar Berisha and Julie Liss, aims to "authenticate the human origin of voice recordings at the point of creation", using sensors already found in many devices and embedding the result as a watermark. This is provenance: proving what is real rather than catching what is fake.

The FTC's own summary is the most useful line in the announcement: the results "highlight that there is no single solution to this problem". The Berkeley team reaches a similar view. They suggest imperceptible watermarks in SI-generated voices, while warning that scammers "would simply migrate to those services that opt out" of such a scheme, and they note that today's forensic techniques mostly work after the fact rather than while a call is unfolding.

Do voice cloning tools stop people cloning someone else?

Often not. On 10 March 2025 Consumer Reports published an assessment of voice cloning products from six companies: Descript, ElevenLabs, Lovo, PlayHT, Resemble AI and Speechify. It found that four of the six fail to take basic steps to stop unauthorised voice cloning, and called on the industry to add a technical mechanism that confirms the speaker's consent before a voice is cloned. Until that check is common, assume any voice you have heard online could be cloned by someone else.

Can SI or Not check a voice?

No, and we prefer to say so plainly. SI or Not analyses images, text, video and music. Our SI music detector looks for the fingerprint that the Suno and Udio song generators leave in whole songs; it does not detect cloned or synthetic voices, and a real recording with a cloned voice on top is not specifically detected. If a suspicious voice message arrives alongside a profile picture or a screenshot, the free SI or Not detector can check that image or text, which is sometimes the weaker part of the story. For the voice itself, the checks below matter more than any score.

How to check a voice you are not sure about

Whether the voice arrives as a call, a voice note or a viral clip, the useful questions are about the request and the source, not the sound.

  1. Notice the pattern. Urgency, secrecy, a crisis, and a payment by wire transfer, crypto or gift cards. That combination is the warning sign, whatever the voice sounds like.
  2. Hang up and call back. The FBI's advice is to verify a caller by hanging up, finding the real contact details yourself and calling that number directly. For a relative, call their usual number or someone who is with them.
  3. Use a family code word. The same FBI announcement recommends creating "a secret word or phrase with your family to verify their identity". Agree it in person, never by message, and do not post it anywhere.
  4. Never pay or share codes on the strength of a voice. No bank, employer or official should need a one-time code, a password or an urgent transfer on a call you did not expect.
  5. For a viral clip, trace the source. Where did the audio first appear? Is there a video, a transcript or an official channel that confirms it? A recording of a public figure saying something shocking should be covered by reliable media if it is real. The same source-first habit applies to how to tell if a video is SI.
  6. Use a detector only as one clue. If you try a voice detector, check what it was built to recognise, and read a "likely real" result as "nothing found", not "proven human".
  7. Limit what you post. The FBI also suggests limiting online content of your voice and image, and keeping social accounts private where you can.

What about deepfake audio of famous people?

Politicians and celebrities are easy to clone because hours of their speech are public. The Berkeley paper opens with a real case: in January 2024, an estimated tens of thousands of Democratic voters in New Hampshire received a robocall in an SI-generated voice of President Biden telling them not to vote in the primary. The authors note that the voice was made with ElevenLabs, the same cloning service they used in their study. For such clips, ask where the audio came from and whether anyone independent has confirmed it. Our explainer what is a deepfake covers how these fakes are made and the cases that have been documented.

The bottom line

A voice is no longer proof of identity. People mistake cloned voices for real ones far too often to rely on hearing, and SI voice detectors, while better than ears in the lab, lose ground on real-world audio and on generators they were not trained for. Use detectors as a clue when you have the file, prefer provenance when it exists, and for anything that asks you to act, verify through a channel the caller does not control: call back, ask for the code word, and take the time a scammer does not want you to take.

FAQ

How can you tell if a voice is AI?

Usually not by ear alone. In a 2025 study in Scientific Reports, listeners identified cloned voices as AI only about 60% of the time. Listen for odd word choices or a flat tone as a first alarm, but confirm by hanging up and calling the person back on a number you already know, or by asking for a family code word.

Is there a free AI voice detector?

Some companies offer free classifiers, but many only recognise their own output. ElevenLabs' free AI speech classifier, for example, estimates whether a clip was made with ElevenLabs and says it cannot detect audio from other providers. A low score from such a tool says nothing about other voice generators.

How accurate are AI voice detectors?

They can score well on the benchmark they were tested on and much worse elsewhere. The ASVspoof 5 challenge, with 53 teams, found that performance degrades under adversarial attacks and after neural encoding or compression. A 2022 study found detectors performed poorly on real-world recordings of celebrities and politicians. Treat any score as one clue.

Can SI or Not detect a cloned voice?

No. SI or Not checks images, text, video and music. The SI music detector looks for the fingerprint of the Suno and Udio song generators and does not detect cloned or synthetic voices.

What should I do if I get a call from a voice I think is cloned?

Do not pay, share codes or act on the call. Hang up, find the real contact details yourself and call back, or reach the person through someone who is with them. A secret family word or phrase, agreed in person, is a simple extra check recommended by the FBI.

Sources

  1. People are poorly equipped to detect AI-powered voice clones · Scientific Reports (Barrington, Cooper, Farid), 31 March 2025
  2. FTC Announces Winners of Voice Cloning Challenge · Federal Trade Commission, 8 April 2024
  3. ASVspoof 5: Evaluation of Spoofing, Deepfake, and Adversarial Attack Detection Using Crowdsourced Speech · IEEE TASLP (arXiv preprint, Wang et al.), 7 January 2026
  4. Does Audio Deepfake Detection Generalize? · Interspeech 2022 (Müller et al.), 30 March 2022
  5. AI Speech Classifier · ElevenLabs, read 11 November 2026
  6. New Report: Do These 6 AI Voice Cloning Companies Do Enough to Prevent Misuse? · Consumer Reports, 10 March 2025
  7. Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA) · FBI Internet Crime Complaint Center, 3 December 2024

Spotted an error? Write to hello@siornot.com. Corrections are made in the article and the update date changes. Read our editorial policy.