In the following column, PPL’s Chief Information Officer Mark Douglas examines the potential for ‘clustering’ to help solve problems ranging from accuracy to fraud…
At a recent London gathering of DDEX, one topic rose above all others: the role that DDEX could play around the grouping or ‘clustering’ of sound recordings.
Despite the somewhat benign description, it is a genuinely interesting subject that leads to several related topics, including the role of
music recognition technology, the nascent International Standard Content Code (ISCC), and even touches on the other hot topic of fraudulent streams.
But before I get into clustering and the substance of the debate, some of you may be asking what DDEX is and why it might be driving such discussions. For those unfamiliar with DDEX, it is the not-for-profit body that develops and promotes the adoption of standards for communicating music data across the entire music value chain.
Created in 2006, DDEX has played a vital role in defining a standard vocabulary and a suite of file formats for exchanging data across our industry. One could argue that our industry only functions today as a result of DDEX.
To get into the heart of the clustering debate, we must first do a short refresher on the International Standard Recording Code (ISRC). Every distinct sound recording is meant to be identified through one, and only one, ISRC. Only when a sound recording has undergone material change, typically through the application of some creative input, should a new ISRC be assigned.
The simple act of re-releasing a recording or including it in a compilation, should not trigger the creation or assignment of a new ISRC. Truly different versions of recordings be they the Radio Edit, Album Version, 12” Remix, Live Version or whatever, should each have their own
These rules are very clearly set out in IFPI’s ISRC Handbook. Yet, despite this clarity, adherence to them has historically been patchy. A search in the IFPI’s ISRC search tool for any evergreen hit of the last few decades shows that successful recordings are typically registered many times over with multiple ISRCs.
When looking for the reasons behind this widespread failure to follow the rules on ISRC assignment, ignorance of the rules, difficulties in determining the existing ISRC or simple human error are all factors. More worrying is that registering a recording under a new ISRC can have economic upsides.
The primary problem that arises is that new ISRCs cause new entries to be made in collecting society repertoire databases. Those databases are used as the basis for allocating large sums of money to the rights owners and performers on sound recordings. When there are multiple distinct entries in those databases, this creates the opportunity to have inconsistent data on the rights ownership and performer line-ups on the different instances of that recording.
“ISCC, if it progresses as planned, has the potential to unlock value for the music industry.”
Whilst it has not always been the case, in recent years, most of these large databases have evolved to correctly de-duplicate recording registrations that have the same ISRC. Such de-duplication allows for conflicting claims of ownership or inconsistent performer line-ups to be surfaced and resolved. However, as identified above, the easiest way to bypass such checks, particularly of ownership, is to register a recording with a new ISRC.
On one level this is fine – someone in the world has released that recording with a different ISRC and that data should be preserved. The new ISRC is likely to be already burned into physical product and held in DSP databases so ignoring it or deleting it would be neither helpful nor efficient.
The problems arise when you look at how money is assigned to rights-holders and performers. When music usage is reported by many types of users (radio, TV, public performance etc) there is often no ISRC data provided. Instead, the music that has been used is described by reference to the track title and band/artist name. In these cases, matching logic must be used to determine which sound recording in the repertoire database is being referred to.
When there are multiple instances of a recording with the same, or very similar, metadata, it can often be a bit of a lottery as to which entry in the database is selected. If the one chosen is an outlying registration with a different ISRC and with incorrect rights owner or incomplete line-up data, then the money flows to the wrong parties or in the wrong amounts.
This is where clustering comes in. And by clustering, I mean the process of grouping together multiple recordings that appear to be identical yet exist under multiple ISRCs. At PPL we refer to it as Recording Roll-Up. Clustering fulfils several important roles. Firstly, it allows patterns of rights ownership across multiple registrations of similar or identical recordings to be assessed. Such an assessment is aided by the fact that ownership of the catalogue of any given band or artist tends to sit with a single rights owner.
Obvious exceptions to this would be the recent re-recording of her studio albums by Taylor Swift, but this tends to be the high-profile exception and not the rule. By having data on recordings that are in a ‘cluster’, a simple data query is all it takes to find unusual claims of ownership across multiple recordings. It is interesting to note that the prevalence of multiple ISRCs is greatest for the highest-value recordings which, I believe, makes the need to do this type of check even greater.
“Clustering is a subject that even touches on the hot topic of fraudulent streams.”
In a similar vein, clustering helps in the maintenance of performer line-ups. It is unfortunately still the case that the performer data included at the point of registration of a sound recording is usually incomplete. This is especially true for non-featured and session musicians. Typically, the line-up data on recordings is made complete through a combination of research by collecting society staff and by the performers themselves reviewing line-ups of songs they know they have contributed to and submitting claims to be added.
When multiple copies of a single song exist, the performer would have to review the line-up of each and every one of them and make an individual claim for each case where they need to be added. At PPL, by using our recording roll-up tools we are able to propagate a single performer addition across multiple instances of a recording where we deem it be appropriate.
For me, the case for clustering is clear. It reduces the risk that the wrong rights owner is paid for a recording, and it assists in making performer line-ups as complete and consistent as possible. The question then turns to how the clusters should be determined. One way to do it, and it is the way we started at PPL, is to use the metadata for each registration and to group recordings with high degrees of similarity in this metadata.
However, care needs to be taken. The simple inclusion or exclusion of vital words or phrases from the recording title, the omission of duration, or the erroneous population of recording date with e.g. release date can introduce error into the clustering process. At a minimum, the fallibilities of recording metadata are such that the logic used to form clusters should err on the prudent side, and a degree of manual review and intervention should be included in the processes that make use of the cluster data.
Recognising the challenges of a metadata-based approach, PPL has been exploring alternative approaches, notably ones that use music recognition technology to verify more reliably which recordings do and do not sound the same. In our case, and based on an assessment of several market offerings, we now make use of recording cluster data from Gracenote. Using their MRT technology, Gracenote are able to de-duplicate sound recordings down to a single, normalised recording, based on their audible similarity.
One of my own test cases for that assessment was around the Kraftwerk catalogue where the 2009 remasters of their studio albums were correctly differentiated from the original 1970s and 1980s releases. The multiple year effort to develop those remasters had created sufficient audible differences to keep them separate from the originals (not to everyone’s liking!).
Whilst this data set has added to the checking that we do at PPL, it is still not the end game. The reality is that there is not one definitive cluster for a given recording. What gets included or excluded from a cluster should be driven by the task that the cluster data is supporting. For performer line-up checks, the rules for inclusion need to be tight; live versions, extended mixes or remixes often include additional or different performers, or other audible contributions such as mixing engineers or studio producers, and those versions should not be clustered.
For a rights ownership check however, it may be entirely appropriate to widen the membership of a cluster. As identified previously, the ownership of any artist’s catalogue often sits with a single rights owner so the inclusion of different versions and mixes can often be appropriate and useful.
In thinking about why such checks are important, it is interesting to note how much focus there currently is on fraudulent streams, bootlegs etc that result in money flowing to the wrong parties. I would contend that there is a much simpler path to getting paid for recordings you do not own – just file some recording registrations with new ISRCs naming you as the rights owner. Why go to the bother of actually making or releasing a new recording?
In the absence of clustering type approaches, there is every chance that usage will match to your version, and you will be paid! There will of course be safety nets to catch large payments going out to unusual or unexpected parties, but these tend only to be effective for a small quantity of very large payments. Clustering supports a vastly improved level of checking at scale.
But clustering is not just of use on the recordings side of the industry. There is a potentially large benefit to be derived by forming much wider clusters spanning multiple bands/artists to, for example, aid the propagation of links between recordings and underlying musical works. But this, and the previous examples, requires flexibility in how clusters get formed. They also require specific technology approaches. MRT solutions in the past have been about finding identical recordings. Finding similar recordings, and being flexible about what ‘similar’ means, calls for new solutions.
One development that may assist in this is the International Standard Content Code (ISCC). At its simplest, ISCC is a fingerprint of digital content. It is designed to work not just on audio files but on any form of digital content. ISCCs do not get assigned or issued. They are calculated by an open-source algorithm that inspects the content of the digital file.
Under the traditional approach to fingerprinting, the fingerprints of two similar but slightly different sound recordings would bear no relationship to each other. Under ISCC, the fingerprints of similar recordings would themselves be similar. In fact, the degree of difference between two fingerprints is indicative of the level of difference between the
Whilst still in its embryonic stages, a solution such as ISCC could form the basis of clustering. By adjusting the tolerance of how far apart any group of fingerprints can be, one can adjust what does or does not get included in a cluster. ISCC is a live project in its early stages but, if it progresses as planned, it has the potential to unlock value for the music industry.
And this is where DDEX enters the frame. If the industry is to start using multiple clustering datasets with a variety of rules about how the cluster has been derived, then is there a role for DDEX to set some standards on how clusters get formed and communicated between parties? Also, who can lead our industry in assessing the viability of the ISCC initiative and help shape the way in which it develops, to ensure our needs are met? These are both topics that DDEX is actively engaging in, and they are topics that are worth keeping a close eye on.
All physical subscribers will receive a complimentary digital edition with each issue.Music Business Worldwide