A 130,000-track dataset for sample identification, and what it can't clear for you

Four researchers, R. Oguz Araz, Xavier Lizarraga, Xavier Serra and Dmitry Bogdanov, have posted a paper on arXiv describing a new public dataset for music sample identification. It is an extended abstract for the late-breaking demo session of ISMIR 2026, first submitted on 14 September and revised on 26 October.

What they built

Sample identification, as the paper defines it, means matching an element of one musical work to the transformed versions of it used in new works: the chopped, pitched and filtered loop, not a clean copy. The authors say the task has had little attention and lacks large public datasets (arXiv).

They mined sampling annotations from a music database and split them into about 114k training tracks, 6k validation and 10k test. They say that is nearly three orders of magnitude larger than existing benchmarks (arXiv).

The interesting work is in the split. A naive split put the same tracks into both training and test sets, so the authors built a graph of who-sampled-whom and split it by connected components. One component held about half of all annotations, which made balanced splits impossible, so they trimmed it (arXiv). That mega-component is no surprise to anyone who knows how a few breaks get sampled again and again.

The dataset is shared for non-commercial scientific research only; the analysis and splitting code is public (arXiv).

Why sample ID matters for money

A sampled record uses two copyrighted works at once. The U.S. Copyright Office treats a sound recording and the musical composition it embodies as two separate works (Circular 56A). Clearing a sample means finding both owners, which starts with knowing what was sampled. Our guide to the two copyrights in every song explains that split, and our explainer on stem separation covers why cleaner extraction tools don't change it.

What we think

This is useful plumbing, not a product. The paper reports no baseline models or accuracy figures, so nobody can yet say how well a system trained on it finds a heavily transformed sample. Treat any claim that "AI can now spot every sample" with suspicion until results appear.

Two limits stand out. First, the labels come from annotations in an existing music database, which can only record samples someone noticed; the obscure, well-hidden ones that cause the nastiest late claims are exactly the ones least likely to be labelled. That is our inference, not a finding of the paper. Second, the non-commercial licence means rights holders and distributors can't simply take this data and ship a scanner, though the method and code can be reused.

The bigger point: if sample detection does get good, it cuts both ways. Producers could check their own tracks before release, and catalogue owners could scan releases at scale for claims. We'd rather producers had that tool first. As with AI-music detectors, a match score should start a conversation about clearance, never settle one.