Using a 16-bit 44khz .wav format file of approx. 8 seconds of silence, I wrote a piece of software that generates a pattern based on the length of the file, and scatters out bits and pieces of the watermark in the file, according to that pattern.
The serial number was 7 digits and was embedded randomly into the file, using a randomized pattern, later stored in a \"key-file\". This way, the only way to decipher the watermark within the sound file, is to use the keyfile which holds the randomized pattern data used in that particular .wav.
I then mixed some music I\'ve done with the watermark datafile. The result is inaudible. And even in mp3-format I\'m able to extract the watermark data from the .wav and read it.
I imagine this must be how it is done. Probably a bit more complicated, but the general idea seems to be the same.
Here are the sound files I used in my test:
(this is the generated digital watermark)
(this is the sound sample with the embedded watermark in - can you hear it?)