Tool 02 — Audio
Is this voice cloned?The free AI voice detector that shows its measurements.
A free AI voice detector for cloned and deepfake audio. A real spectral analysis runs in your browser: high-frequency band limits typical of neural vocoders, channel facts, encoder trails and AI tool fingerprints in the metadata.
Aipurity is a free AI voice detector for cloned and deepfake audio — it shows you the recording’s measured frequency spectrum, not a black-box score, and nothing leaves your browser.
- Runs 100% in your browser
- Voice audio is never uploaded
- Shows the measured spectrum, not a vibe
Drop a file, or click to choose
Analysed locally in your browser. Nothing is uploaded, ever.
What this AI voice detector actually reads.
Spectral band-limit analysis
A real FFT runs in your browser. Neural vocoders and TTS typically render at 16–24 kHz, leaving a hard ceiling far below what a real microphone records.
Channel forensics
Cloned voices are usually generated mono and duplicated to stereo. We compare channels sample by sample and flag identical pairs.
Encoder & toolchain trail
FFmpeg versions, encoder tags and container metadata reveal how many times a file has been transcoded — and what wrote it last.
AI voice fingerprints
ElevenLabs, Suno, Udio, Resemble, Murf, PlayHT, Descript and others sometimes leave their names in metadata. When they do, the verdict is immediate.
How to check if a voice is AI-generated.
- 01
Decode locally
Your recording is decoded by the browser’s own audio engine. Nothing is uploaded — voice data shouldn’t go to a random detection site.
- 02
Measure, don’t vibe
Six Hann-windowed FFT frames are averaged; we report the actual frequency where energy stops and show you the spectrum.
- 03
Signals with caveats attached
A 16 kHz ceiling can mean a vocoder — or heavy MP3 compression. Every signal says what else could explain it.
How voice cloning actually works
Modern voice cloning is neural text-to-speech: a model turns text into a mel-spectrogram, and a vocoder renders that spectrogram into audio. Three seconds of your voice is enough to train a passable clone with today’s tools. The vocoder step is where physics leaves a trace — most render at 16–24 kHz sample rates, producing a hard frequency ceiling that a real microphone recording doesn’t have. That ceiling is measurable — the one thing an AI audio detector can actually stand on — and this check measures it instead of asking you to trust a score.
Shown evidence vs. a black-box score
Most AI voice detectors return a single percentage. Here’s the practical difference when the answer matters:
| Aipurity | Typical score-based detector | |
|---|---|---|
| What you get | The measured spectrum, channel facts, encoder trail and any named fingerprints | One number, e.g. “87% AI” |
| Can you verify it? | Yes — the evidence is displayed, caveats attached | No — the model’s reasoning is opaque |
| Compressed audio | Flagged as a caveat: MP3 truncation explained next to the spectral reading | Silently skews the score |
| Where your audio goes | Nowhere — decoded in your browser, works offline | Uploaded to the vendor’s servers |
Who runs this check
Families
A voice note that doesn’t sound quite right. Check the file’s evidence before you act — and verify anything urgent with a callback on a number you already have.
Newsrooms & fact-checkers
A “leaked” recording of a public figure. The spectral ceiling, encoder history and fingerprints are checkable evidence you can cite, not a vibe.
Small businesses
A voicemail from a “supplier” or the “CEO” authorizing something unusual. Run the file, then verify out-of-band — vishing runs on urgency.
Musicians & labels
A track that might be a Suno or Udio export of your voice or style — their names in the metadata settle it instantly.
Recruiters & HR
Submitted audio introductions and “interview” clips can be synthetic; band-limits and channel duplication are quick first checks.
The 60-second checklist
| Check | Why it matters | Then what |
|---|---|---|
| Spectral ceiling at 16–24 kHz | Neural vocoders render band-limited audio; real microphones don’t stop there | Weigh against compression — the tool shows both readings |
| Identical stereo channels | Clones are generated mono and duplicated; real stereo recordings differ sample-to-sample | Strong supporting signal when the source claims to be a live recording |
| Encoder & toolchain trail | Fresh FFmpeg tags with no recorder metadata mean the file was rebuilt | Ask where the “original” actually came from |
| AI tool fingerprints | ElevenLabs, Suno, Udio and others sometimes name themselves in metadata | A named maker is conclusive; absence proves nothing |
| Out-of-band verification | The best clone can’t answer a callback on a known number | For anything involving money or safety, always do this last step |
The phone call that isn’t your family
Cloned voices are the engine behind scam patterns that exploded since 2024:
The family-emergency call
“Grandma, I’m in trouble — send money now.” A cloned voice plus urgency. Defense: hang up and call the person back on the number you already have. No detector replaces that callback.
Business vishing
A “CEO” or “supplier” voice authorizes an urgent transfer. Fraud reports tracked deepfake-driven fraud up over 1,000% in North America in early 2025. Verify out-of-band, always.
Leaked audio & blackmail
“We have a recording of you saying this.” Synthetic blackmail audio is now cheap to produce. Before panic or payment, check what the file itself admits — and preserve the original for evidence.
The interview that never happened
Fabricated podcast clips and “hot mic” moments spread faster than corrections. The encoder trail often shows a rebuilt file with no recording chain behind it.
Where this check helps
Received a voice note, a “leaked” recording, or a suspicious voicemail file? Run it here before you share it or act on it — metadata fingerprints and spectral ceilings survive in original files.
The scale of voice fraud
Three figures from the sources cited below — the reason this check exists:
+1,100% in one quarter
The measured jump in deepfake-driven fraud attempts in North America in early 2025, per Sumsub’s identity-fraud reporting.
$16.6B reported lost
Total cybercrime losses reported to the FBI’s IC3 for 2024 — voice-enabled social engineering is a growing slice of it.
8 million deepfakes forecast for 2025
Up from roughly 500,000 shared in 2023 — a 16× rise in two years, per figures collected in the Sumsub report.
Voice cloning, in plain words
The terms behind the measurements on this page:
Neural TTS & voice cloning
Text-to-speech trained to imitate a specific voice. Modern systems need seconds of reference audio, not hours — which is why any public voice can be cloned.
Mel-spectrogram
The intermediate “picture of sound” a TTS model generates before audio exists. The vocoder then paints it into a waveform.
Vocoder
The neural component that renders spectrograms into audio. Most render at 16–24 kHz sample rates — the physical origin of the frequency ceiling we measure.
Band limit / frequency ceiling
The frequency where a recording’s energy stops. Real microphones capture content past 20 kHz; vocoders and lossy codecs both cut off earlier — which is why this signal always ships with a caveat.
Channel duplication
Synthetic voices are generated mono; “stereo” exports are often two identical copies. Real stereo recordings differ between channels at the sample level.
Voice conversion
Reshaping a real performance into another person’s voice — the hardest case for any detector, because the timing and emotion are genuinely human. Out-of-band verification is the defense.
Formats & limits
| Source format | What happens to the highs | Evidence quality |
|---|---|---|
| WAV / FLAC (lossless) | Full band preserved to ~22 kHz+ | Strongest — a vocoder ceiling stands out unambiguously |
| M4A / AAC | Moderate high-frequency shaping | Good — ceilings remain readable with caveats |
| MP3 (low bitrate) | Hard truncation of highs by the codec itself | Weak for spectral reads — metadata and channel checks still work |
| Phone calls / WhatsApp voice notes | Band-limited by the network and app to begin with | Spectrally blind — fingerprints occasionally survive; verify out-of-band |
Anything your browser can decode works — MP3, WAV, M4A/AAC, FLAC, OGG, WebM — and there’s no size cap because there’s no upload.
Honest limits
What it can’t tell you.
Voice-clone fraud rose more than tenfold by 2025 industry reports — and the best clones, laundered through a phone call, leave no spectral smoking gun. For anything high-stakes (a “relative” asking for money), the defense is a callback on a known number, not any detector.
Common questions.
The spectrum cuts off at 16 kHz. Is it fake?+
Maybe — or it’s just an MP3. Lossy codecs and neural vocoders both truncate highs. That’s why this signal alone never produces a “detected” verdict.
Can it tell which tool cloned a voice — ElevenLabs, Murf, PlayHT?+
Only when the tool says so itself: some exports carry the maker’s name in metadata, and we read ElevenLabs, Suno, Udio, Resemble, Murf, PlayHT and other fingerprints. A clip laundered through a call, a screen recording or a platform re-encode loses that trail.
How can I tell if a voice recording is AI-generated?+
Run it through the checker above — free, in your browser. Look for a hard spectral ceiling at 16–24 kHz, identical stereo channels, a suspicious encoder trail, or an AI tool’s name in the metadata. Any single signal has innocent explanations, which is why we show all of them with caveats instead of a single verdict.
How is this different from ElevenLabs’ own AI Speech Classifier?+
ElevenLabs’ classifier uploads your audio to their servers and is built to recognize ElevenLabs-generated speech. This check runs entirely in your browser, reads evidence from any tool that leaves it, and shows you the measured spectrum instead of a single score.
How accurate are AI voice detectors?+
Honest answer: nobody has proven high in-the-wild accuracy, and black-box scores are easy to over-trust. We report measurable facts — spectral ceilings, channel duplication, named fingerprints — with the alternative explanations attached, so you know exactly how strong the evidence is.
Can you detect a phone-call deepfake?+
Telephone audio is band-limited by the network itself, destroying the spectral evidence. For personal safety, verify out-of-band: hang up and call back.
What file formats are supported?+
Anything your browser can decode — MP3, WAV, M4A/AAC, FLAC, OGG, WebM. There’s no upload cap because there’s no upload.
Does it work on a phone?+
Yes — the FFT and metadata parsing run in mobile Safari and Chrome exactly as on desktop. Save a suspicious voice note to Files and open it here; it never leaves the device.
Can it detect AI music — a Suno or Udio song?+
Often, yes: music generators are among the most reliable metadata-stampers we see, and a Suno or Udio name in the file ends the question. A re-encoded rip from a streaming platform loses that trail, and the spectral signals are less meaningful for music than for speech — so expect “detected” on originals and “inconclusive” on laundered copies.
Does it work for languages other than English?+
Yes — the physics doesn’t care about language. Spectral ceilings, channel duplication, encoder trails and metadata fingerprints are language-independent signals, so a Mandarin or Spanish clip reads the same way an English one does.
What if the check is wrong?+
Single signals can mislead — that’s exactly why we never turn one into a percentage. A band limit might be compression; pristine audio might be a high-end clone. Treat “detected” (a named tool in metadata) as reliable, treat everything else as weighted evidence, and make high-stakes decisions with an out-of-band callback, never on any detector alone.
Is my voice data uploaded?+
Never. The FFT and metadata parsing run in your tab; the page works offline once loaded.
Sources & further reading