Allegro

Watermarking and Fingerprinting Explained

Volume 125, No. 4April, 2025

Ken Hatfield

This month’s submission for Local 802’s A.I. series is by Ken Hatfield, a member of Local 802 since 1977.

The digital paradigm shift with its ethos of “move fast and break things” has left many of us complacent. Disruptive aspects of tech innovations have become so commonplace that we accept them without question because they provide convenience. But what is the true cost of this convenience?

Tech workers often refer to themselves as disrupters. They are proud of moving fast and breaking things. Disruptive uncertainty is now a fact of everyday life. About all we can count on from the tech world is that there will be perpetual change, that change will be disruptive, and it will all be driven by the pursuit of profits.

This is having devastating consequences for the music business, consequences that are about to get far worse because of the existential threat AI poses. With AI we face the most disruptive technology to date… one that can and will replace humans…one that not only threatens our livelihoods but our identities as musicians.

Musicians need protection from the unbridled use of our work to train AI to replace us! These protections must compensate ALL musicians on recordings used to train AI, because it’s not just featured artists that make a recording what it is. The contributions of sidemen often comprise distinctive aspects that are characteristic of many recordings, aspects that AI will digest in its training and use in its self-generative capacity.

What would Paul Simon’s “50 Ways to Leave Your Lover” be without Steve Gadd’s iconic drum parts? What would Bill Withers’ “Just the Two of Us” sound like without Marcus Miller’s famous bass line that opens the track? Try to imagine Marvin Gaye’s “What’s Going On” without James Jamerson’s bass parts. The list of tracks shaped by the contributions of sidemen is endless. Training AI to create “facsimile” alternative replacement recordings (which are cheaper to produce) will inevitably involve digesting and copying more than just the compositions/songs or the featured artist contributions. It will require digesting and copying what the sidemen played. When sidemen’s work generates revenue due to its use to train generative AI to create audio tracks which will be monetized, shouldn’t their contributions on the recordings that trained AI be compensated too? The current crop of pending legislative protections only compensates featured artists and rights owners (labels, publishers, and composers).

To address these inequities, the Local 802 Artificial Intelligence Advisory Committee and the Jazz Advisory Committee of Local 802 have proposed a revision to recording contracts. Please note that this proposed revision is a first step. Legislative actions will also be required. But as a primary stakeholder group that will have to deal with AI, the musicians union needs to be on record demonstrating what protection(s) we require. Here’s our proposal:

“The Local 802 Artificial Intelligence Advisory Committee and the Jazz Advisory Committee of Local 802 have proposed that all union recording contracts include a requirement that all recordings produced under such contracts employ watermarking technology so that all participants and copyright/right owners on the recording may be identified.”

Without comprehensive, accessible metadata identifying all participants on recordings, tech will simply claim they cannot identify any but the featured artists, even if we get legislation requiring sidemen compensation. This will result in only rights owners and featured artists being compensated.

Like most businesses, tech has limited options for increasing profits. Among the methods they favor, automation wherever and whenever possible dominates their business practices. For the protections we seek to be viable, they must be compatible with the automated systems tech relies on. For digital audio, such systems are called “automated audio recognition”.

So far, two main approaches for automated audio recognition have emerged in the marketplace: active (watermarking) and passive (fingerprinting).

Watermarking is an active recognition method because it augments, or watermarks, audio files by adding metadata to the zeroes and ones that comprise a digital audio file. Much like the long-established method of adding a visible watermark to paper, an audio watermark is an identifier…a unique electronic identifier embedded in an audio file/signal. It is readable via software, yet inaudible to the listener.

Digital watermarking embeds information (metadata) into a digital file/signal (e.g., audio, video, or pictures) in a way that is difficult to remove. If the file is copied, the metadata is also carried in the copy.

Embedding metadata via watermarking in digital audio files occurs during the mastering process: after recording and mixing, but before manufacture and/or commercial release. When an audio watermark is added to an audio file, a recognition system can then access, identify, and read all metadata embedded within that digital audio file. You can include any information you choose (such as featured artists, sidemen, composers, publishers, and rights owners, even ISRC#s for each track of a CD). This metadata can be used to identify ALL participants and rights owners on a digital audio recording.

Watermarks can easily be detected by a variety of software applications. This means that tech can use a wide array of software based automated recognition methods to access the embedded meta data, which makes watermarking a cost-effective means for tech to identify all participants on recordings used to train AI. And we want ALL participants on such recordings to be identifiable so they can be compensated.

Most studios include watermarking with the mastering process for no additional charge. Many software companies making recording software include watermarking in their suite of options for no additional fee. The fees that will compensate sidemen in our proposal won’t come from record labels or publishers. Those fees will come from the tech companies that used our recordings to train AI without our advance consent.

Tracks recorded before the advent of watermarking can either be remastered adding the watermarked metadata retroactively or fingerprinted via the passive option referenced above.

Fingerprinting is a form of passive recognition because it requires no modification of the source material. Fingerprinting recognizes an audio clip by comparing it to a reference database and determining whether a match can be found. Audio fingerprinting is based on the idea that every audio recording is unique. And just like a fingerprint, every recording’s unique characteristics can be used to identify it.

There are various types of fingerprinting. They are all essentially different methods of generating a unique identifier that can be used to identify a digital audio recording. Here are some of them:

Spectral fingerprinting: which is based on analyzing the audio’s frequency spectrum to generate a unique identifier.
Temporal fingerprinting: which is based on analyzing the audio’s temporal structure to generate a unique identifier.
Hybrid fingerprinting: which combines both spectral and temporal fingerprinting methods to generate a more robust unique identifier.

For digital audio files to be “recognizable” by fingerprinting technology, they must be registered in at least one of several databases. These databases are generally owned and operated by businesses that offer services including audibly recognizing and identifying songs. If you’ve ever used Shazam to identify a song, you’ve used fingerprinting. Here’s a partial list of companies offering fingerprinting: Shazam, Audible Magic, Gracenote, PEX, BMAT.

You simply go to one of these companies online and register recordings of your songs (which requires uploading audio files of your songs). When you register you can include the metadata you choose for each recorded song. That makes the recorded songs (and all the info you include) accessible so entities needing to pay all participants on such recordings can identify who is eligible for compensation.

So, we have a variety of methods for assuring we will be identified and ultimately compensated for our work.

There is another aspect of this that is salient for all sidemen. Back in the heyday of recording when folks bought physical product, it was commonplace with vinyl and later CDs to include liner notes and credits with the packaging. The credits identified all the musicians on recordings. That created work for musicians. If you played on a recording and people dug what you played, being recognizable meant other musicians and producers could identify you and hire you for their projects. Well, all the audio formats used by Digital Service Providers (DSPs) for things like streaming, including lossless formats like FLAC files, are based on the MP3. That means there are very limited metadata fields available. Consequently, sidemen routinely go unidentified. Having the expanded metadata both watermarking and fingerprinting can accommodate increases the likelihood that your work will not only be noticed, but that you will be acknowledged. That acknowledgement is going to generate more work for you.

By the way, tech is notorious for ignoring the limited metadata they currently receive. So, we may need legislation making such behavior unlawful.

In future articles we will address legislative and judicial protections that will be needed. But as Chinese philosopher Lao Tzu famously said: “The journey of a thousand miles begins with a single step.”

Our proposal for the revision of AFM recording contracts is that first step for professional musicians. So please actively support it.

If you are interested in participating in future AI committee meetings, please send a request to be notified of upcoming meetings along with your email address to Local 802 Recording Vice President Harvey Mars at hmars@local802afm.org.

Ken Hatfield has been a member of Local 802 since 1977. Send feedback on Local 802’s A.I. series to Allegro@Local802afm.org.