CORPUS Journey
CORPUS began with a question: what if the musicians owned the library that machines learn from? Part of the answer, the world is still not ready for. The other part became a bridge to the industries that need music.
In Lure, an artistic research project with Volkswagen's machine learning lab Argmax, we explore what learning machines can do with instruments, voices and narration on stage. We keep running into the same wall: there is almost no good, well-described music for machines to learn from, let alone in the quantity needed.
Could musicians build a shared library of music together, on fair terms, owned and shaped by the people who make it? The idea is bigger than anything we could build at the time, so we keep it in a drawer.
ChatGPT arrives, and generative AI is suddenly everywhere. Music follows within months. The question we put in a drawer is now everyone's question.
At the end of the year, the New York Times sues OpenAI, and training data moves to the centre of the debate. The idea stays with us, and we begin to share it with friends.
A Creative Europe Innovation Lab call gives the idea its moment. We write it all down: interactive access to sound and music, built on a library that musicians contribute to on fair terms. CORPUS starts in December 2024, initiated by Sofilab. Now we speak for the project, with musicians, lawyers and institutions such as GEMA.
We build the integrity checks, the rich semantic descriptions and the contribution app. We take our proposal to the music industry, a fair way to license music for AI training. The reactions are sceptical, but we believe a good idea can win them over. In October we present CORPUS in Munich, with live music and an open discussion.
The answer arrives. Musicians and rights holders want their music heard, used and paid for in films, series and games, and most do not want to license it for training machines. Our proposal was to license music for training with the musicians’ consent, before any model learns from it. But at the end of 2025, the major labels settled with the AI companies that had trained on their music without asking, and licensed them after the fact. Next to deals like that, an offer to do it properly from the start found few takers.
Along the way, we have built something else: integrity checks and rich semantic descriptions for every track, so it can be found and licensed with confidence. Filmmakers tell us they cannot find music with character, and that licensing for sync takes far too long. CORPUS becomes the bridge. Soundbook opens as a public beta, and catalogues reach chat assistants through MCP. Training our own models stays possible, for those who choose it.
Once the rights chain is in place, with permissions and ownership for every track, we leave the beta. From the first search to the signed licence, productions, catalogues and musicians work along one path. Soundbook becomes the place where film, series and game makers find their music.
With every track described in depth, Sofilab builds the first proofs of concept for products that listen: a score that follows the player, driving atmospheres that respond to the road, a bedside device that sings along. The music comes from the corpus, and the musicians are paid. Later, music can follow a player, a drive or a patient as the situation changes, made with musicians who choose to take part.
The question from 2021 never went away. Every licence pays the musicians. But the lasting value of a platform grows in the organisation itself, and once the corpus is large and trusted, the musicians who built it should own a share of it. Many musics, one corpus.
Be Part of What Comes Next
CORPUS is built by musicians, sound engineers and musicologists, together with everyone who brings their music in. If this journey speaks to you, there is a place for you in it.