Freesound Forums

Hi all,

for some time I've been curious about digital watermarks in sounds ( http://en.wikipedia.org/wiki/Digital_watermarking ), so I decided to do some research. Obviously I have no intention of marking any sounds that I release, that would be pointless under a CC license... probably (?). My interest on the issue was plain curiosity... along with a wish to get some knowledge on a process that advocates of copyright see as a solution to many of their problems. Here I 'll share my - very limited - knowledge with you.

You can read the Wikipedia article which is clarifying, but in a nutshell one can create a watermark within any sound by mixing it with a pattern of hiss buit from any image. That's cool because you can use any common jpg file, from a picture of your pet to any text or logo.

In the repositories of Linux I found the free program named Enscribe (by Jason Downer, http://www.coppercloudmusic.com/enscribe/). Very easy to install and (although command line only) easy to use. Equivalent programs for windows may exist or not, I don't know.

I really don't know (or care) how the conversion of image into sound is made, I'm not so techie. Just guess the process uses lots of FFT transformations, or whatever, What matters to me is that pixels are converted into noise (hiss) which distributes along the whole range of audible frequencies following the same pattern shown by pixels in the photograph.

Alternatively, the hiss may be created using just part of the frequency range. For example, you may decide that only frequencies above 14000 Hz will be used in the output file. In this case the resulting sound will be barely audible, yet will be there for anyone to see under a spectral view.

If one then blend (mix) the sound file containing hiss only (in a pattern) with another sound sample (like a song of your favourite rock band) the resulting file will be a song indelibly marked. You cannot hear the hiss, yet the song has a stamp on it. A stamp made of hiss. What amazed me most at first was that the size of the un-stamped and the watermarked songs were identical. How could this be? But it's quite simple if you take into account the fact that file size depends basically on the duration of the sample, NOT on its content (and duration is not changed after adding hiss, obviously)

I finally put what I had learnt into practice, made some watermarks, watermarked a couple files, and uploaded to Freesound as pack (http://freesound.iua.upf.edu/packsViewSingle.php?id=3153). You may check it.

An interestingly corollary is: if you watermark at a high frequency (above 12000 Hz, for example), but then compress the sample - to ogg or mp3- chances are that the watermark will disappear completely because compression often gets rid of high frequencies. If on the other hand, you put the stamp inside the range of audible frequencies, the quality of the sound will be affected (the hiss may be heard). Could the pattern of noise be somehow removed in this case? I don't know,

Saludos

D

http://www.postimage.org/aV1RY8J9.jpg

nice.

But in general I cannot know if an audio file is watermarked or not, right? I would think that especially in the case of fingerprinting any pictorial patterns are concealed?

-afaict

afaict

But in general I cannot know if an audio file is watermarked or not, right? I would think that especially in the case of fingerprinting any pictorial patterns are concealed?

-afaict

afacict, I guess that will depend on how 'secret' you want the mark to be. On occassions one might want the mark to be visible/audible in order to discourage copying...
But even if the mark was completely hidden, a bit-by-bit comparison of a tagged file vs a non-tagged one would reveal the difference, or so I guess.

D

well...

if you've ever opened a small jpg or other image file in notepad, what you get is a bunch of unintelligible, seemingly "random" symbols. I'm not sure how it is decided which symbols mean what in the audio file, but if you open a bitmap as raw data in audacity, you get a short burst of noise, i'm assuming there's some correlation between the symbols in notepad and the noise you hear when its opened as raw data. Seeing as how in the text file the symbols seem almost random, we all know that noise is an unrepeating waveform that is randomly offsetted

ive been doing experimentations with corrupting short sound files using wordpad's "replace" function, and also been creating text files with only single ASCII characters repeating to see what they sound like, i'll upload some to freesound this weekend maybe

hope i contributed something that provided at least a bit of insight.

I've actually run into this once before, or something along the same lines, when tools 10,000 days album came out a bunch of my friends and I got together and started tearing it apart looking for secret messages and so forth. Well we didn't find anything but one of the interesting tidbits I did come across was these pictures in the sound. Turns out it's been going on since about 99 here's a link to one of the sites I found. It's got links to programs that help do this kind of thing...
any way enjoy

Link:
http://www.bastwood.com/aphex.php

dobroide
But even if the mark was completely hidden, a bit-by-bit comparison of a tagged file vs a non-tagged one would reveal the difference, or so I guess.
D

could we have the same in sound?
does exist a tool that calculate the difference between two sound files giving the result as a sound file, instead of those inconprehensibobble esadecimal symbols?
where there is no difference there will be silence - so the difference is audible.

if it's possible to have such result, then the Author could retain his original sound file and publish the wavemarked file -> the difference brings out the audible wavemark.
(watermark = water = wave = wavemark)

Interesting stuff, Dobroide.

Something that occurs to me as a potential problem would be what happens if two watermarked files are mixed together ? There may be a possibility of the high frequency components beating together and producing audible erroneous lower frequency differences. e.g. if one watermark involved a frequency of 15 KHz and another 16 KHz at the same time offset, then a difference note of 1 Khz would appear in the mix.

@lysander, nice link, beautiful images, thank you!
@cajo, dunno if the editing soft I use has this feature (substracting one sound files from another), I'll check

and @acclivity, I catch what you say (albeit intuitively, technicalities surpass me). In fact I tried to do something similar with the mosquito alert sound ( http://freesound.iua.upf.edu/forum/viewtopic.php?t=2804&start=0 ) without much success...

cheers

you guys should look at:

http://www.google.be/search?q=robust+watermarking+audio

there are plenty of papers written on the subject, so look through the proceedings of ICMC ( http://www.icmcXXXX.org where XXXX is a year ) and ISMIR ( http://www.ismir.net/ )

Some watermarking schemes are (somewhat) robust to compression like mp3, mixing and distortion, but of course none is perfect.

You should also look at "fingerprinting" which is kind of the inverse. You fingerprint audio (your sample) and then later you can recognize the audio again in another context.

- bram

You need the tool Coagula, it allows you to load a picture and it converts it to a wav file.

The file below was created in that tool:
http://freesound.iua.upf.edu/samplesViewSingle.php?id=31921

^^^ Looks like an interesting program.

Cajo asked if there were any programs that can find the difference between two audio signals. The answer is yes, any multi-track audio editor worth it's salt can do that (Audacity is open source, hint hint).

Just line the two files up, invert one of them and then mix the two together. The result is (signal A) - (signal B). If the two signals are audibly identical, but one has a watermark, the final product of mixing those two signals will be the watermark by itself, surrounded by silence.

You do have to be careful to line the files up properly though. I don't know of any automated way to do that. Remember that different formats may add or subtract a handful of leading samples at the beginning of a file, so you've got to zoom way in to be sure you're lined up.

it works easy and quick in Audacity! thanks a lot alatham82

the idea is to add a wavemark/watermark in a sound file. here's my home made experiment - no expertise, no programming, only two afternoons wasted.

I thought of Morse code because it's builded on time intervals, not on frequency.
used WinMorse to build a sample to insert http://www.winmorse.com/
the text is "freesound", here in dots and lines: ..-. .-. . . ... --- ..- -. -..
and in wav format: http://www.freesound.org/samplesViewSingle.php?id=56297

the concept is: silence = original sound / Morse character = altered sound.
if the alteration is very light it would be inaudible, thus the wavemark is hidden.

to face the most difficult situation, I worked on the perfect sound of a Tibetan singing bowl, shared by hanstimm:
http://www.freesound.org/samplesViewSingle.php?id=15361

here's the marked sample (wavemark begins 5sec after start):
http://www.freesound.org/samplesViewSingle.php?id=56357

to check the result I used technique suggested by alatham82, inverted one sample then mixed with the other: the result is an audible wavemark
http://www.freesound.org/samplesViewSingle.php?id=56358

comparing original and marked converted in mp3 (128kbps 44khz), the wavemark is still audible.

found no sw to translate an audio file back to Morse. need some seawolf to check wavemark intelligibility