Cloned-voice scams often arrive as a short voice note: a relative in trouble, a boss who needs a payment sent today. You can run a note like that through a detector — but voice notes are among the hardest audio there is to judge, so how you get the file out and how you read the result matter as much as the check itself.
A detector answers whether the audio looks synthetic. It cannot tell you whether the request is genuine, and it cannot replace the one check that settles that: contact the person yourself, on a number or account you already had, not the one that sent the note. If you cannot reach them, ask a question only they could answer, or use a code word your family or team agreed in advance.
Run the detector as well if you want a record, or if the note is evidence you need to assess — for a report to your bank, your employer, or the platform. Just do not wait on it before making that call.
Detectors need a file. These are the formats the common apps hand you, and whether TextSight accepts them directly.
| Where the note is | What you usually get | Accepted as is? |
|---|---|---|
| WhatsApp voice note (share or export the message) | An .opus file (Opus audio) | No — convert to WAV or MP3 first |
| Telegram voice message (save or forward to files) | Usually an .ogg file | Yes, as OGG |
| iPhone Voice Memos (share) | An .m4a file | Yes, as M4A |
| iMessage audio message (save to Files) | Often a .caf file | No — convert to WAV or M4A first |
| Voicemail (export from your phone or carrier app) | Varies by phone and carrier | Check the extension; convert anything other than MP3, WAV, M4A, OGG or FLAC |
Apps change their export behaviour over time, so trust the file extension you actually end up with over this table. If the file name ends in something the detector does not list, convert it.
Convert once, from the original file, to WAV: it adds no further compression on top of what the app already applied. Any audio converter will do. If you have ffmpeg installed, this is the whole job:
ffmpeg -i voice-note.opus voice-note.wavAvoid the shortcuts that seem easier: playing the note on one phone and recording it on another, or screen-recording the chat. Both add a second microphone, a room and another round of compression, and all of that pushes the result around.
Messaging apps compress voice heavily to save data, and that compression removes much of the fine detail detectors use. The consequence runs in an uncomfortable direction: genuine voice notes can come back leaning synthetic. A high AI score on a voice note is therefore much weaker evidence than the same score on a clean recording.
The fix is a baseline. Find an older voice note from the same person, in the same chat, about something unremarkable — same app, same phone, same compression. Convert it the same way and run it first. If the old note scores clearly human and the new one clearly synthetic, that difference means something. If both lean the same way, the compression is driving the result and the audio cannot settle the question. The reference-clip method has the full reading table.
Any one of these is a reason to pause and call back, whatever a detector says. Report the account to the platform, and if money moved, contact your bank immediately. For organisations, the fraud teams guide covers payment controls that do not depend on recognising a voice at all.
Your first check is free with no signup — 3 a day with a free account. Your audio is processed and discarded, never stored.