To hide digital data in recorded music in such a way that it is difficult to deliberately remove, resistant to re-recording via a loudspeaker and microphone, resistant to MPEG compression, and inaudible to trained ears, while still being easily recoverable by an appropriately equipped forensic engineer.
To write the watermark, a pattern is added (to record a zero) or subtracted from (to record a one) the spectrogram of a section of the music. The pattern is chosen to minimise audibility of the resulting change and to be unique to the individual applying it and difficult to guess.
Before reading the watermark, Bayesian inference was used to construct a probability distribution for recorded music of a sufficiently high standard that it can distinguish the following three objects:
- Original music;
- Music with added pattern, then distorted, then subtracted pattern;
- Music with added pattern, then distorted, then with the pattern added again.
To read the music, Bayesian inference is again applied to distinguish these three constructs.
The ability to prove that music contains data identifying its originator is primarily useful for providing forensic evidence of copyright infringement. Other potential uses include the covert transmission of information (steganography) and (where the reading algorithm and keys are made available to the listener) automatically directing the listener to a particular website or other resource.