Audio Watermarker

● Embedded on your device with an open AI model; nothing is uploaded.

Loading…

Hide an inaudible ID inside an audio file with AudioSeal, Meta’s open watermarking model, running in your browser. Name each copy before you send it — one per label, client or reviewer — and if a copy turns up somewhere it should not, the checker reads its ID back and tells you whose it was.

How it works

  1. Drop the audio and name each copy you want: one recipient, or a list of them.
  2. Download the watermarked files and the receipt that maps each ID to its name.
  3. Later, drop any copy (and your receipts) on the checker to read its ID.

Questions

Does my file get uploaded?
No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
Will people hear it?
It is designed not to be heard: the mark sits far below the music, and it follows the loudness of the audio rather than adding a steady hiss. Listen to the result in the page before you send it; on very quiet spoken passages a careful listener with headphones may notice a faint texture.
What survives, and what removes it?
In our tests the ID came back exactly after MP3 at 64 and 128 kbps, AAC, a messenger-style voice-note compression, resampling, a phone-band filter, turning the volume down and room echo. A clip cut out of a marked file is still detected as marked, but its ID can occasionally be misread, because the ID is sensitive to exactly where the cut falls. Changing the pitch or the speed removes the mark, and heavy added noise weakens it. It is a strong deterrent and a tracing tool, not an unbreakable seal.
Does it prove I own the audio?
No, and no public watermark can. The tool is free for everyone, so anyone can stamp any ID onto any file. What it reliably tells you is which of YOUR copies a file came from, which is what leak-tracing needs. It also only finds its own mark: it is not a detector for AI-generated audio.
How many different IDs are there?
The mark carries 16 bits, so 65,536 IDs, written like WM-3F2A. Each copy gets a random one unless you type your own, and the receipt file keeps the name you gave it next to the ID.

Related tools