Hacker News Reader

Comment by dTal | Hacker News Reader

>I don't think there is anything inherently wrong with watermarking. The problem is with this particular poorly tested implementation.

There is. Audio watermarking is inherently audible, because that is the only way to avoid being stripped out (deliberately or inadvertently) by the incredible technology we have for removing inaudible information from audio recordings: lossy audio codecs. The artifacts must be so salient that the codecs deem them important, audible information - codecs designed with elaborate psychoacoustic models, especially to make that judgement call with maximum efficacy. There's no way out.

View on HN

If the listener doesn't know the "true" information, it is sufficient if the watermarked version sounds plausibly like a real sound. I think it's too strong to require the watermark to be completely inaudible. It is sufficient if given two recordings of unknown sounds a listener can't tell which one is watermarked and which one isn't.

As a musician; no. That is not sufficient. Any audible watermark -- and as the grandparent comment correctly points out, that is any that is going to do its job -- is unacceptable. It is modifying what the artist intended you to hear in a destructive way. It is destroying the original performances. This is a terrible practice.

Are you against all lossy compression methods as well? The watermark can simply be a slight tweak of the compression algorithm that could as well have been done for non-watermarking reasons.

I think the point is that the watermark is supposed to survive being re-encoded, so it can’t be something just in the compression algorithm.

There is a discussion to be had around watermarks surviving transcoding but I was mostly trying to understand whether the opinion

> It is modifying what the artist intended you to hear in a destructive way. It is destroying the original performances.

also applies to lossy encoding.

The whole point of lossy compression is that it tries to be inaudible, as a primary concern.

The whole point of audio watermarking is that it survives lossy compression. Inaudibility is merely a secondary concern.

This is an adversarial relationship, and one that the audio watermarking is doomed to lose. By and large, lossy codecs succeed in transparency, and therefore as an inevitable consequence, watermarking fails at it. After decades of development, lossy codecs are just too good - there's nowhere left to hide information.

(It's also worth noting that the consumer benefits from the quality tradeoff that compression makes, in the form of decreased storage and bandwidth. They don't benefit from the audio watermarking at all.)

There are always going to be some subjective decisions in the lossy encoding process. Determining which bits to drop via encoder configurations/customization is far from an exact science. These decisions themselves can act as a watermark.

At sufficiently high bitrates modern lossy codecs are transparent. Even MP3, a dinosaur of a lossy codec is transparent (as in nobody can successfully ABX it) for a lot of music starting somewhere in the neighborhood of 192kbps, and at the 320kbps that commercial stores sell MP3s it is transparent for all physically generated music so far tested (some electronic music with incredibly fast attacks can have audible artifacts).

No, as was already explained upthread, "the watermark is supposed to survive being re-encoded, so it can’t be something just in the compression algorithm". It needs to be something that survives no matter which subjective decisions the lossy encoder makes.

Graphical watermarks also destroy the original image, and I think are much more intrusive than what an audible one would be.

I'm unable to load the article and haven't heard one in real life, but in theory it could be done in a way that is imperceptible to human ears but detectable by a program. E.g. imagine masking specific frequencies within a noisy section of the video.

There should be less destructive ways for studios to do this, but unless it's in the video or audio signal it doesn't stand a chance of working reliably.

If a signal is imperceptible to human ears but detectable by a program, it's a prime candidate for being thrown away by lossy compression to save space. All the low hanging fruit there has long since been grabbed. The whole point of an audio watermark is that it must survive re-encoding, but it's a race that cannot be won.

The difference is that the graphical watermark on e.g. a photograph usually isn't there when you've paid for it. These audio watermarks are very obvious to me and frankly make instruments sounds out of tune -- it's like the beating you can hear when tuning strings. I basically only listen to classical music and will not tolerate this.

Right, but it's a problem with this specific implementation, not something that couldn't be addressed by user friendlier versions.

To be clear, I'm also against any kind of watermark, but I'd rather it be done with an imperceptible (to me) audio signal than with a constant graphical watermark as is usually the case. Actually, graphical watermarks could be smarter and less obtrusive as well.

> The difference is that the graphical watermark on e.g. a photograph usually isn't there when you've paid for it.

The same could be the case for these audio ones. It's user hostile if this is done on already paid content (but again, I can't read the original article to confirm).

You are thinking about blatant watermarks on stock imagery and the like. Movies at the theater[1] and streaming videos[2] absolutely contain subtle visible watermarks even though you are paying to see them, which is directly analogous to having audible watermarks on Spotify. In the case of movie theaters we are somewhat desensitized to them, as old projection technologies used similar flashing dots for other reasons, like a cue to the projectionist to get the next reel ready[3].

[1] https://creativepro.com/major-movie-studios-specify-digital-...

[2] https://partnerhelp.netflixstudios.com/hc/en-us/articles/440...

[3]https://en.wikipedia.org/wiki/Cue_mark

Oops, that second link is about another type of watermarking, not the anti-piracy kind. Unfortunately, I won't have time to dig up a more appropriate link today.

All kinds of things between the bits on my harddrive and my brain modify what the artist intended me to hear. Is it also unacceptable to listen to songs on headphones, or in rooms with a different geometry than where the song was recorded?

That bar is far too low - reductio ad absurdem, you could go as far as changing lyrics - or substituting an entirely different song - and if you'd never heard the song before you'd never know the difference.

As for more subtle changes, what constitutes a "plausible sound" in any given context is an AI-complete problem. Sure on a particular song, you might get away with e.g. tinkering with the reverb a bit, or changing the tempo by 0.5% - but how can you possibly do that in the general case?

Remember - all such changes have to be audibly salient to humans, because that's the only thing guaranteed to survive.

The connections made between artist expression and listeners are precious, rare, vital.

You’re saying it’s ok to decimate it - as long it’s replaced with some “plausible” proxy? Plausible decided by who, the artist or business requirements?

The argument is that this exists on a spectrum. The best way to get the artists true intention is to watch a film in a well-calibrated theater. Then a high-quality Blu-ray. Web streaming involves even more compromise.

We tolerate Netflix releasing films in less than 4k. They do that because of a practical business benefit (cost). Similarly, Universal is making a compromise in the artist's vision to meet a business requirement. I think if the compromise is sufficiently small and the business benefit is sufficiently large, it's ok to do this.

If I piss in your cereal, and dont tell you, is it still ok if it tastes semi ok?

As a former orchestra player, and a listener who likes to sit in front of a Hi-Fi system, and listen to music intently, this is not OK.

When you're passive listening, you're only perceiving the "outlines" of the music, but listening actively reveals a lot of details, and this kind of wobbling becomes both more noticeable and disturbing.

When listening the music actively, anyone can distinguish between natural and unnatural sounds, and this will bother many people if they listen more actively.

I imagine this could be worked around by using forked encoders that are designed to introduce the watermark when encoding lossy versions.

This of course would mean the watermark would not survive transcoding from lossless, but as we can see in TFA the watermark can be made useless all the same without such efforts.

With that said the oxygen malpractitioners that run the music industry have never put actual consumers' satisfaction above implementing trivially circumvented security theater, so pirates win all the same in all scenarios.

Huh? Couldn't anyone just re-encode the lossy version? Though with some generation loss, since a transparent lossy encoding presumably would also preserve any watermarks.

Yes. The entire supposed value proposition of watermarks is to survive re-encoding / the analog hole.

I disagree. It depends how many bits you want to encode in what period of time. If you only need 1 bit per second, then it can still be inaudible and not susceptible to _current_ lossy psychoacoustic compression models. Universal was both dumb and greedy (and hopeful?) so they got caught making their music into crap.

However, encoding a series of highly discernable tones into the audio is just dumb. Instead they should use phase modulation at low frequencies. Current compression models don't play with subtle changes of meter much, because they mostly use local encoding and there's not much room for improved compression with (non-audible) modulation. Of course, if Universal wanted to be able to accurately detect 5 second clips (rather than 2 minute songs) this would be a lot more difficult... and of course they want every thing, at the expense of listener experience.

All major market FM broadcasts in the United States have audio watermarking applied to them, and most likely to their online only streams as well. Do you hear the watermarking in that audio?

The system is called CBET and it's based on the psychoacoustic model as well, except it determines the level at which watermark should be injected at. High enough that it passes codecs easily, but just below perception so you generally can't detect the tones that it has added to the audio.

The trade is that CBET only moves about 8 bits/second of information through the side channel, which is just enough to push a station ID and timestamp through the channel for Nielsen's ratings purposes.

FM radio is a dumpster fire anyway, quality-wise. Apart from the awful dynamic range and noise floor, god knows how many plugins and mp3 re-encodes are used in the average pipeline. So I wouldn't be especially surprised to find that 8 bits/second extra crap on the pile goes unnoticed.

However I'm awfully curious about this system, especially "high enough that it passes codecs easily". I would naively expect that to be highly dependent on the codec and bitrate chosen, and further expect the range between "inaudible" and "codec-passing" to be negative for many common untransparent lossy profiles, like 128kbps mp3. Where can I read more?

It's hard to find information on. The company used to be called "Arbitron" and they filed several patents on the technology, which is where I did most of my research.

There are 10 bands, each of which can carry 1 of 18 tones. 16 tones are to signal, so the stream is 4 bits wide, and the other two tones are used as a "STOP" and "SYNC" marker in the data stream.

All bands carry identical information, but each bit within a band is encoded with a different tone than any other band. So, it appears random or uncorrelated at first glance, but you can see the pattern pretty quickly if you do a long enough analysis. The same message gets identically repeated 12 to 13 times per minute and the message only changes once per minute. Here's an example isolation of the signal [1].

The coder monitors incoming audio and does a psychoacoustic pass to determine which level the tones should be injected at. If it can't find a good level for a tone, it uses the lowest possible level... which usually works out fine due to the 10 redundant signals in the watermark.

Encoders like MP3 see the injected tones as signal and will make the bits available to ensure they're encoded rather than correctly masking them out as non-audible material. The 10 bands help here as well, as even if the encoder fails to code the signal in one band, there's usually enough redundant bands that you can successfully decode the watermark.

[1]: https://www.sigidwiki.com/wiki/CBET

That CBET signal is oddly alluring. Must kill...

There's a large gap between audible and detectable.

A transmitter could transmit a weak signal in a specific pattern on a number of arbitrary preselected frequencies. They would more-or-less seem like noise lost in the noise floor. Only when you specifically filter for the specific set of frequencies can you discern the pattern that was put there that could not appear by chance. You don't even need to use the same frequencies throughout the transmission. As long as the schedule is decided up-front it can serve like a one-time pad for encryption.

> Audio watermarking is inherently audible, because that is the only way to avoid being stripped out (deliberately or inadvertently) by the incredible technology we have for removing inaudible information from audio recordings: lossy audio codecs.

You're assuming that perceptual audio coders do a perfect job at encoding what can be heard* and not encoding what can't be heard*. This isn't true in both directions — they encode detail that can't be heard* in the source, and lose detail that can be heard in the source*.

*This is contextual, depending heavily on the listener and the listening conditions.

Not a perfect job. But a state of the art job! The relationship between watermarkers and lossy perceptual codecs is fundamentally adversarial - one side trying to eliminate all inaudible information, the other trying to sneak inaudible information past it. And so much more research has been poured into lossy compression, that the watermarkers just can't win. Bear in mind also that the watermark needs to work with all lossy codecs, so the information must be considered salient by all of them. There's just no way to do this inaudibly.

I doubt that this is where we are at today. We have higher fidelity transmission compared to a joint-stereo 128kbps mp3, which was probably already enough to carry imperceptible watermarks.

I am far from an expert in this, but, couldn't you design a watermark that removes certain frequencies in known locations in the spectrum, at known times? Those gaps would obviously stay missing in a compressed version - lossy codecs don't add noise.

I didn't say _perceptual_ watermarking, which, as you point out, is at odds with perceptual audio compression... :)

Reply on news.ycombinator.com