GEMA launches PLAI, a licensed AI training dataset, as Klangio signs on first

GEMA has launched its first licensed dataset for training AI music tools. Called PLAI by GEMA, it sells cleared music to AI companies so they can train on licensed content instead of scraped catalogues. The German collecting society named Klangio, a transcription firm from Karlsruhe, as its first customer.

The pitch is one-stop rights. PLAI bundles the authors’ rights to the works, the master rights to the recordings, the sound files, and the metadata from a single source. For an AI developer, that means one license instead of chasing publishers and labels separately.

Five production-music partners supply the catalogue: Bailer Music Publishing, Earmotion Audio Creation, Intervox Production Music, Music Sculptor, and Sonoton Music. GEMA says the offering can grow as more publishers and collecting societies join, which is how it plans to scale the dataset for different customers and use cases.

What PLAI by GEMA gives AI developers

The dataset ships with heavy metadata, not audio alone. Each track carries genre, tempo, key, instruments, and chords, plus segment timestamps, ISWC and ISRC codes, structure data, and mood and context tags for sync use. Training packages can be tailored to a specific model and use case.

On payment, GEMA says every contributing rights holder is compensated proportionally. It frames PLAI as a fair, licensed alternative for AI providers who want to train on cleared content. What it has not published is the actual split: how a license fee gets divided among the publishers and authors, or how the value of any one track is measured.

Why some in the industry call it too small

Not everyone reads the launch as a breakthrough. Matthias Strobel Hohmann, president of MusicTech Germany, applauded the move in principle but questioned both the timing and the scale.

“

An AI model trained solely on stock and library music does not learn songwriting, emotions, or the stylistic diversity of popular music. 178,000 audio files is a joke.

His point is about scale. A generative model that writes full songs trains on millions of tracks with real vocals and arrangements, so a curated slice of production music will not teach it to write. He also called the revenue model a black box, since license fees appear to flow to a handful of publishers with no public formula for the split.

Klangio’s answer: PLAI is not built for song generators

Sebastian Murgul, co-founder and CEO of Klangio, answered the critique head on. His company is the first PLAI customer, and he agreed that 178,000 tracks would not move the needle for a song generator. That, he said, is the point.

“

There's a whole world of music AI beyond generation, transcription, analysis, education, MIR, where you simply don't need millions of tracks. Training on professional, fully licensed music is a real benefit for us.

Klangio’s tools turn audio into sheet music and MIDI. A model learning to read notes and chords needs clean, well-labeled, licensed music more than it needs raw scale, so a curated set of production music fits it. Murgul called PLAI a first step in the right direction.

The timing, one week before the Suno verdict

PLAI arrives one week before the Munich Regional Court rules in GEMA’s case against Suno on July 31, 2026. GEMA is suing over exactly this problem, AI trained on protected songs without a license. Launching a licensed product days before the verdict lets GEMA show the court a working alternative.

For independent artists, the launch matters less as a product and more as a marker. It is the first time a major collecting society has put a cleared, paid dataset in front of AI developers, following the licensing model GEMA proposed in 2024.