The idea
Every person who listens to or downloads a track here gets a copy that sounds identical to everyone else's, but carries a hidden, inaudible mark that is unique to them. If a copy later shows up somewhere it shouldn't, you can upload it to the Identify page and Tracemark tells you whose copy it was.
It is a deterrent and a piece of evidence, not copy protection: nothing stops someone from copying a file, but they can no longer do it anonymously.
How the mark is made
The audio is cut into short segments. For each segment the server prepares, once, two almost identical versions: A and B. They differ by a very faint, noise-like pattern that is shaped to hide under the loudness of the music or speech itself, so you cannot hear it.
Each user has an ID. When you play or download a track, the server builds your copy by picking A or B for every segment according to the bits of your ID. The result is a unique sequence that is repeated through the whole track. Because only two layers are stored, adding a new user does not mean storing another copy of every file.
How a leak is identified
Detection is blind: you only need the suspect file, not the original. The server decodes it, looks for the secret patterns in each segment, works out for every segment whether A or B was used, and reads the user ID back from that sequence. Because the ID is repeated, damage to some segments can be corrected by combining the others. The result comes with a confidence level, and files that have no mark are reported as "not detected" rather than guessed. The patterns depend on a secret key that stays on the server, so without it the mark cannot be found or forged.
What it survives (and what it doesn't)
These results come from our own test suite: 4 clips (speech, piano, synthetic pop) x 4 users, with every copy re-processed the way a pirate might.
| Treatment of the copy | Outcome |
|---|---|
| MP3 128 kbps and 64 kbps, AAC 96 kbps, Opus 48 kbps | User identified every time |
| Chains of codecs (e.g. MP3, then AAC) | Identified every time |
| Resampling (48 to 44.1 kHz, down to 22 kHz), volume change, low-pass at 4 kHz | Identified every time |
| Trimming the start and end, combined with volume change and re-encoding | Identified every time |
| Mild added noise (30 dB signal-to-noise) | Identified every time |
| Very low bitrate (MP3 32 kbps) | Weaker: about 3 of 4 identified |
| Heavy added noise (20 dB signal-to-noise) | Weaker: about 4 of 5 identified |
| Speeding up or slowing down by even about 1% | Defeated: mark no longer found |
| Averaging or mixing copies from several users | Defeated or unreliable (collusion) |
| Very short excerpts (under about 20 seconds) | Not enough audio to read the ID |
In our checks, unmarked files and files tested with a wrong key were never mistaken for a marked copy. These numbers come from a small test set and are a guide, not a guarantee. Re-recording the sound through speakers is also not covered.