This is a report on the state of vocal data licensing in 2026. It is written for ML teams, legal counsel, investors, and commercial operators trying to understand where the market is, how it got here, and what is coming next. The data is drawn from primary legal sources, publicly disclosed commercial deals, and direct observation of the market from inside The Vocal Market's enterprise program.
If you read only one section, read the summary below. If you have time for more, the sections after it provide the context and specifics behind each of the headline findings.
Summary: what changed in 2025 and what 2026 looks like
The most important development in the last twelve months is that the question "should we license training data or scrape it?" has been definitively answered by the courts and by the major labels. Two headline events drove this:
- The $1.5 billion Bartz v. Anthropic settlement, preliminarily approved September 25, 2025, became the largest copyright settlement in U.S. history and established a concrete price tag for training on pirated data.
- The UMG-Udio and Warner-Suno settlements (October and November 2025) moved the major label strategy from litigation to licensing partnerships, with artist opt-in frameworks and joint AI music platforms.
The result is that in 2026, the supply of licensed training data is growing, the demand for licensed training data is outpacing supply, and the legal risk of relying on unlicensed data has been priced to a level that is no longer rational for any company building to scale.
For the vocal data subcategory specifically, the picture is more nuanced. Vocal data is structurally scarcer than instrumental data because the consent chain for voice is more complex (performer rights, moral rights, right of publicity). The market for licensed vocal data lags behind the market for licensed instrumental data by roughly 18 months. We expect 2026 to be the year that lag closes.
Section 1: The legal weather
The current legal environment for AI training data is the most unfriendly to scrapers it has ever been. Three ruling patterns combined to produce this.
Courts are rejecting fair use for commercial AI training
Thomson Reuters v. Ross Intelligence (D. Del., February 11, 2025) became the first federal court ruling to reject an AI fair use defense. Judge Stephanos Bibas ruled that Factor 4 (market harm) weighed decisively against the AI defendant, calling it "undoubtedly the single most important element of fair use." The ruling involved non-generative AI but the market-substitution logic carries over directly to generative cases.
Bartz v. Anthropic (N.D. Cal., June 23, 2025) produced a split summary judgment from Judge William Alsup: training on legally purchased books was fair use, training on pirated books was not. The two-track ruling established that fair use is not a blanket shield and that the provenance of training data matters in the fair use analysis.
Kadrey v. Meta (N.D. Cal., June 25, 2025) was a procedural win for Meta, but Judge Vince Chhabria explicitly wrote that plaintiffs making a properly developed market-dilution argument "very well might have won." The opinion is widely read as a roadmap for future plaintiffs, not a green light for AI defendants.
The settlements are getting bigger
The $1.5 billion Bartz settlement is the headline number, but the Suno, Udio, and Stability AI cases are collectively on track to deliver hundreds of millions more in aggregate settlement value over the next 12-24 months.
Warner Music settled with Suno on November 25, 2025, becoming the first major label to strike a licensing partnership with an AI music generator. Universal Music settled with Udio on October 29, 2025, establishing a joint AI music platform with opt-in artist compensation. Sony Music continues to litigate against both defendants as of April 2026.
Europe is moving faster than the U.S.
The Munich Regional Court ruled against OpenAI in November 2025 in a case brought by GEMA, Germany's music collecting society, over the reproduction of copyrighted song lyrics in ChatGPT outputs. It was the first worldwide copyright ruling against OpenAI and it was decided in Europe. OpenAI is appealing. The ruling is a strong signal that European courts will take plaintiff-friendly positions on AI copyright questions more readily than U.S. courts.
The EU AI Act's General-Purpose AI provisions became applicable on August 2, 2025, with enforcement beginning August 2, 2026. Article 53(1)(d) requires GPAI providers to publish a "sufficiently detailed summary" of their training data. Penalties reach €15 million or 3% of global turnover. The mandatory Training Data Summary Template published on July 24, 2025 formalizes the disclosure requirements.
Section 2: The market for licensed music data
Three events in 2024-2025 established that licensed training data is a viable commercial supply chain at scale.
Stable Audio 2.0 proved the licensed-only model
Stability AI's Stable Audio 2.0 was trained exclusively on licensed data from AudioSparx, covering over 800,000 audio files with artist opt-out honored. It is the cleanest "we trained only on licensed data" public claim in the space, and it became the reference point against which other commercial AI music systems are now measured.
Meta MusicGen proved the mixed-source licensed model
Meta's MusicGen trained on approximately 20,000 hours of licensed music, combining internal Meta tracks with Shutterstock and Pond5 content. The model is released under CC BY-NC for non-commercial use, but the underlying licensing framework is a template for commercial systems that are now following.
The major labels went from litigation to partnership
The Warner-Suno and UMG-Udio settlements changed the relationship between major labels and AI music generators. Before the settlements, the labels were plaintiffs. After, they are commercial partners with licensing fees, equity stakes, and joint product roadmaps. The joint AI music platform UMG and Udio are building in 2026 is a direct product of this shift.
This has implications for the data market. Major label catalogs, which were previously unavailable at any price, are now licensable at high prices and under specific commercial structures. The practical effect is that the top tier of the training data market is becoming accessible to companies at scale, while the middle and lower tiers (stock libraries, dedicated datasets) are where most training data flows.
Section 3: The vocal data subcategory
Vocal data has distinctive market characteristics that make it different from instrumental data.
Performer rights add a layer
Instrumental recordings involve session musicians, producers, and labels. Vocal recordings involve all of those plus the performer's distinctive right in their voice. In the EU, performers hold neighboring rights under Directive 2006/115/EC independent of the sound recording copyright. In the U.S., the ELVIS Act (Tennessee, 2024) added voice to right-of-publicity protection, and the federal NO FAKES Act is pending.
The practical effect is that a vocal dataset needs to clear rights at two layers (recording + performer) where an instrumental dataset only needs to clear one (recording). This makes vocal datasets harder to assemble at scale and creates a natural supply constraint.
Public data is structurally insufficient
As of 2026, the total pool of open-source clean singing data is approximately 230 hours, heavily weighted toward Mandarin Chinese. English clean singing data at scale essentially does not exist in open repositories. Most open singing datasets (OpenCpop, OpenSinger, M4Singer, GTSinger) are released under non-commercial licenses.
This gap between what commercial teams need (hundreds to thousands of hours of English and multilingual singing data) and what open data provides is the structural reason for commercial vocal dataset licensing to exist as a category.



