The algorithms

We now present how we think the marking and the detection algorithms work. It should be underlined that these are simply suppositions derived from a very limited amount of material. However, these suppositions seem to fit rather well on the three songs that were provided in the challenge.

Let us now briefly explain why we believe the detection algorithm works this way. The embedded mark is very small. It is actually a noise compared to the signal of the song. The standard technique to detect a noise embedded in a signal is correlation. However, one needs to correlate on a long enough chunk so that the noise correlation is much larger than the correlation of the signal and the noise.

Consequently, correlating on $1470$ samples is not enough to reveal the presence of the mark. This is why we are actually correlating on the average of $p$ chunks of $1470$ samples.

We have tested the detection algorithm with two different sizes of $p$, $p=30$ (one correlation per second) and $p=450$ (one correlation every $15$ seconds, the maximum detection time required by the original SDMI call for proposals). The results are given in section 6.