AudioContext fingerprint

The reading is silent by design. The sound is rendered into memory rather than sent anywhere, so there is nothing to hear and nothing to allow.

1. 1

   Renders a tone offline

   A tone runs through a compressor in a context that never reaches the speakers.

2. 2

   Reads the samples back

   The rendered channel comes back as an array of numbers the page reads directly.

3. 3

   Adds a slice of them up

   That total is the fingerprint, and on an untouched machine it comes out the same every time.

The number comes out of the sound stack rather than out of a setting. The compressor implementation, the rounding of very small values and the build all leave their mark on it.

That makes it coarse on its own and strong in company. It survives with nothing stored on your machine, so clearing storage does not move it and neither does a new window. A site needs one render and one sum, and that takes a fraction of a second.

Adding the same small amount everywhere looks like protection and reads like a mark. One subtraction against a reference recovers both the amount and the places it was added.

A fixed step makes it worse. A mask learned from one sound then fits every other sound the page renders, and the defence becomes the trait that identifies the profile.

The shift is worked out for each sample from the profile seed, the position in the buffer and the bits of the sample itself. There is no step to find, the amount differs from place to place, and a mask taken off one sound does not fit another.

Silence is left alone, because a shifted nothing is visible. Reading the same buffer twice does not shift it again, so the total a page reads stays put however often it asks. The same work covers the other road out of the sound stack, the one that hands over frequency data rather than samples.

• The shift is applied at all

  The same tone rendered with no seed and with one gives two different totals.

• Reading twice does not move it

  Two totals taken one after the other in one profile come back identical.

• There is no common step

  The distances between shifted samples do not pile up on one value, which is what a period would look like.

• There is no single amount

  The differences at the shifted places are not one constant repeated down the buffer.

The self-test lives in the app and runs inside a profile. Its audio line renders the tone twice and reports the total together with whether the repeat matched.

The free check on this site shows a value for the browser you are reading this in. It is computed on the page and no request goes out with it.

Run the self-test in two profiles on one computer. Two totals where there is one sound stack is the reading worth having. The acceptance run takes the same reading in two profiles before a build ships and refuses the build if both come back with one total.