GRAI, the AI music startup building a social streaming service, has partnered with Songtradr’s MassiveMusic on two deals: one to supply the catalog behind its upcoming app and another to license a dataset for training its own music model, according to a press release dated Oct. 15.
Under the first agreement, MassiveMusic delivers commercial repertoire through a single feed, giving GRAI a route to participating independent and distributor catalogs under MassiveMusic’s rightsholder agreements. Approved titles are available on GRAI within minutes of MassiveMusic sending them, and rightsholders will be paid and receive usage reports for plays.
MassiveMusic also supplies metadata describing tracks across more than 30 dimensions, from mood and energy to instrumentation, key, and era, which GRAI uses for search and recommendations so new listeners can find music before their friends are on the platform.
The second agreement covers training. GRAI says it trains its model on a dedicated, opt-in dataset approved by rightsholders specifically for AI training through MassiveMusic’s opt-in system, and that the model is built in-house to understand the building blocks of a song (melody, harmony, and rhythm) rather than just generate audio. The training license does not itself authorize downstream creative uses of recordings or compositions.
“For us, establishing ‘rights before code’ from day one was an operational necessity,” said GRAI CTO and co-founder Dzmitry Kamarouski.
The deal extends a run of licensing moves. In August, GRAI became one of the first two companies, alongside Udio, to license patents from UMG-backed Music IP Holdings covering the use of recorded music with AI. In September, it acquired Hangout, the social listening platform from the Turntable.fm creators, which launched with licenses from Sony Music, Universal Music Group, Warner Music Group, and Merlin.
Early builds of GRAI are in testing on a test catalog from MassiveMusic ahead of a public launch, and a waitlist is open at grai.fm. The release does not say when the service will launch, which labels or distributors are in the feed, how many rightsholders opted into the training dataset, or when the model will be available.