AI Voice Detector: Can You Spot a Cloned Voice? (2026)

An AI voice detector is a tool that estimates whether a recording was generated by AI. Some detectors read a watermark the generator embedded, and others use a classifier trained on known fakes. Independent tests show accuracy drops on new generators and compressed audio, and listeners are not reliable judges by ear. Treat any result as one clue, and verify the speaker another way.

Can people hear the difference?

Listening studies give a clear answer: not reliably.

| Study | Setup | Result | |---|---|---| | Lavan et al., PLOS One, published 2025-09-24 | 150 participants; human voices from a public speech corpus and voice clones made with ElevenLabs | Labelled human: real voices 62% and clones 58% in one experiment, a difference that was not significant. Real voices 72% and clones 70% in a second experiment. | | Mueller and Choong, arXiv preprint, submitted 2026-05-21 | Large-scale study of human perception of audio deepfakes | Accuracy on real recordings fell from 72.7% in a 2021 baseline to 64.1%. On fake samples it moved from 72.9% to 71.2%. |

Generic AI voices, which were not clones of real people, were labelled human only 41% and 39% of the time, so most of them were spotted. The clones were the problem. The Mueller and Choong result also shows a second effect: listeners grew more suspicious of real speech, which means a person can be wrong in both directions.

How do AI voice detectors work?

There are two main approaches.

Watermark checkers look for a mark that a generator placed in the audio on purpose. A hit is strong evidence for that generator, but a miss says nothing about audio from other sources.
Classifiers are statistical models trained to separate real speech from synthetic speech. ElevenLabs describes its AI Speech Classifier as statistical, and it warns that the classifier "may produce false positives or miss modified audio".

Vendors often combine the two. ElevenLabs' Audio Detector checks for a watermark first and uses the classifier only when none is found.

Watermark checkers vs classifiers

| Tool | Approach | What it can say | Reported limits | |---|---|---|---| | ElevenLabs Audio Detector | Watermark first, then the AI Speech Classifier | Whether audio carries an ElevenLabs watermark, or the classifier's view if not | Only checks ElevenLabs-generated audio. Sign-in required, free on all plans. | | Google SynthID Detector | Reads SynthID watermarks | Whether a file was made with Google or partner AI, with OpenAI, NVIDIA and Kakao named as partners | Binary result. Secondary reports say English only, and that Suno and Udio songs return not detected. | | Resemble AI Detect | Vendor detection model | Company-reported accuracy of 99.47% on audio for its DETECT-World model | Audio detection costs $0.035 per second on the Flex plan and $0.015 per second on Team and Business. No test protocol disclosed. | | Pindrop Pulse | Vendor deepfake detection | Vendor-reported detection rates from short audio segments | Figures published in February 2024. No independent evaluation found. |

How accurate are AI voice detectors?

Vendor claims and independent results are far apart, and no single percentage describes every detector.

Vendor claims. Pindrop's February 2024 article claims 99% detection for known deepfake engines, over 90% for new or unseen engines, and a false positive rate below 1%, from just 2 seconds of speech. Resemble AI reported in August 2026 that DETECT-World reaches 99.47% accuracy on audio. Neither company disclosed its test protocol in the sources we checked.
Classifier figures. ElevenLabs reports 99% precision and 80% recall for its AI Speech Classifier on unmodified ElevenLabs audio. Those figures do not carry over to modified audio or to other vendors' output.
Independent benchmark. Deepfake-Eval-2024 (arXiv 2503.02857) gathered 56.5 hours of audio from 88 websites in 52 languages, all circulated in 2024. Compared with earlier benchmarks, open-source detectors' AUC fell by 48% for audio. Commercial detectors did better, but still did not match human forensic analysts.

The gap between the vendor claims and the independent benchmark is the point. Even Pindrop's own page separates known engines from unseen ones, and the independent benchmark shows that real-world audio is harder than lab conditions.

Why detectors fail

New generators. A detector trained on known generators can miss new ones. Pindrop's own page admits the gap between known and unseen engines.
Compression. The ASVspoof 5 overview, with 53 teams, found that detector performance degrades under adversarial attacks and under neural encoding and compression. Many calls and messaging-app clips are heavily compressed.
Modified audio. ElevenLabs warns that its classifier may miss modified audio.
Short clips. Pindrop claims that two seconds is enough. We found no independent test of how short clips perform.

How to check a clip

Run the clip through the watermark checker for the tool you suspect.
If there is no mark, record that as unknown, not as human.
Use a classifier as a second opinion, and remember that it can produce both false positives and misses.
Check the speaker another way. The FBI advises calling back on a number you find yourself, and agreeing a secret word with family.

A detector result never proves that a voice is real.

What to do instead of trusting one score

The FBI's advice does not depend on detection tools. It says to verify the caller by finding a number for them independently and calling back, and to create a secret word with family members. The FTC advises asking a question only the real person would know. These checks still work when a detector gives no clear answer.

What this means for creators

If you publish AI speech, labelling it is the simplest honest step. Rules are starting to put marking and detection duties on providers. From 2026-08-02, EU AI Act Article 50(2) requires providers to mark synthetic audio so that it is detectable. Under California's AI Transparency Act, as amended by AB 853, covered providers with more than one million monthly users must offer a free AI detection tool from the same date.

This is general information, not legal advice.

OndaLabs does not offer a detector, and this post makes no claim about marks in OndaLabs output. OndaLabs makes speech, dialogue and custom voices from recordings you have permission to use. Its speech to text feature transcribes audio, and it is not a voice detector. Custom voices are available in the voices section.

For how watermarks and provenance data work, read AI audio watermarking. For scam patterns, read voice cloning scams.

Frequently asked questions

Can you tell if a voice is AI-generated by ear?

Not reliably. In a 2025 PLOS One study, listeners labelled ElevenLabs clones as human at rates close to those for real voices. A 2026 preprint found that accuracy on real recordings fell from 72.7% to 64.1%, which means listeners also doubted genuine speech.

Are there free AI voice detectors?

Some are. ElevenLabs' Audio Detector is free for signed-in users on all plans, but it only checks ElevenLabs audio. Google's SynthID Detector was reported as opened to the public in October 2026, and it checks SynthID marks. Each free tool covers a narrow set of sources.

How accurate are AI voice detectors?

No single figure applies. Vendor claims such as 99% come without a disclosed test protocol, while an independent 2025 benchmark found large drops on real-world audio. Accuracy also depends on the generator, the compression and any editing.

Can a detector prove that a voice is real?

No. A detector can find a mark or estimate that audio looks synthetic, but a result saying a voice is not AI is not proof that the person said it. Use a second channel, such as a callback on a number you already trust.

What if a detector gives an unclear result?

Treat it as unknown, not as proof either way. Check the speaker through another channel, and keep the clip for your records.