A voice note arrives, it sounds like someone you know, and something about it is off. This page covers the whole job: exporting the file from WhatsApp on iOS or Android, converting the Opus audio into a format a detector can actually read, and — importantly — how much weight to put on the answer, because voice notes are compressed in exactly the way that makes detection harder.
Two things about the WhatsApp format change how you do this, and most guides skip both.
WhatsApp records voice notes using the Opus codec, normally in an Ogg container, and exports them with a .opus extension. Opus was chosen because it produces tiny files at speech bitrates, which is exactly right for sending a message over mobile data and exactly wrong for audio forensics.
.opus file is not on that list, so it will be rejected. Convert it to MP3 or M4A first — the conversion step below takes about a minute. Renaming the file to .mp3 does not work; the audio has to be genuinely re-encoded.
AI voice detection leans heavily on fine spectral detail, the noise floor between words, and small irregularities in breath and timing. A voice-note-bitrate Opus encode is designed to throw away precisely the parts of the signal a listener will not consciously miss — which overlaps almost exactly with the parts a detector wants.
The practical consequence: on a WhatsApp voice note you should expect lower confidence than on a clean recording of the same speech, and you should read an uncertain result as genuinely uncertain rather than as a soft yes. A confident result on a voice note is still worth taking seriously. An ambiguous one means the file did not carry enough signal to decide, and no amount of re-running will change that.
Save the original file. Do not screen-record the playback or hold one phone up to another — that adds a whole re-recording layer and will wreck the result.
WhatsApp/Media/WhatsApp Voice Notes/ — many devices keep received notes there.Any converter that genuinely re-encodes the audio will do. Here are the routes that do not cost anything.
VLC is free, open source, and handles Opus natively. Open VLC, go to Media → Convert / Save (on Mac, File → Convert / Stream), add your .opus file, choose an MP3 or Audio - MP4 profile, pick a destination and save. This runs entirely on your own machine, so the file never goes anywhere.
One command, no interface: ffmpeg -i note.opus note.mp3. Same benefit — nothing leaves your computer.
Plenty of audio-converter apps on the App Store and Play Store will convert Opus to MP3 or M4A. Prefer one that converts on-device. If the app uploads your audio to its own servers to do the conversion, you have handed a sensitive recording to a third party you know nothing about — which rather defeats the purpose.
Converting cannot restore detail that Opus already discarded, but a low-bitrate re-encode on top will lose a little more. If your converter offers a bitrate setting, choose something generous — 192 kbps or higher for MP3. File size is not a concern here: the detector accepts up to 10 MB, and a converted voice note is normally a fraction of that.
Open the AI Voice Detector, drop in your converted MP3 or M4A, and you get an AI-likelihood score and a human-versus-synthetic verdict in a few seconds. Three checks a day, no signup needed.
| What you see | What it means | What to do |
|---|---|---|
| Confidently synthetic | The audio carries strong signs of machine-generated speech, and enough of them survived the compression to be clear | Treat the message as untrusted. Verify the sender through a completely separate channel before acting on anything it asked for. |
| Confidently human | The recording carries the physical irregularities of a real microphone capture | Reassuring, but not the whole picture. A real recording can still be taken out of context or edited. It says the voice was not synthesised, not that the message is honest. |
| Uncertain | Common on voice notes, and usually a statement about the file rather than about the voice | Do not read it either way. Fall back on verification: call the person on a number you look up yourself. |
Most people checking a voice note are not doing research — they are deciding whether to send someone money. Here is the short version.
Worth saying plainly: the callback step works whether or not a detector is available, whether or not the file converted cleanly, and whether or not the result was confident. If you only do one thing on this page, do that one.
The live-call version of this problem: what to do in the first sixty seconds, and where to report it.
What to do now →Run a known-real voice note from the same person alongside the suspect one. On compressed audio this is the strongest method you have.
Learn the method →Why compression matters so much, what false positives look like, and why we do not publish a single accuracy number.
See our position →Three checks a day, free, no signup. Your audio is never stored.