Starmirror made training an AI choir into a public performance

In Starmirror, the instruction was simple enough to follow by ear. Professional singers offered a phrase; a choir and visitors repeated it. Recording those exchanges turned a public singing session into material for an AI choir.[1]

Holly Herndon and Mat Dryhurst’s project moved from Berlin’s KW Institute for Contemporary Art to Düsseldorf’s K21, where its run ended on 11 October 2026. The exhibition combined a listening environment with the making of new vocal recordings. The act of supplying the model was part of the artwork.[2]

A score becomes a voice

The Ur-Hildegard Training Corpus began with Ordo Virtutum, Hildegard von Bingen’s twelfth-century morality play. A model trained on its music generated a new songbook. KW credits Valentin Crogiez and Mathieu Giraud of the Algomus research group with the model research and generation, and Thomas Fournil with musicology and arrangement.[1]

That sequence gives people several creative roles. Someone has to interpret the historical notation, prepare the new score and perform it. The resulting recording becomes material for a further model. What the next system learns therefore carries decisions made at each stage, including how singers turn written instructions into sound.[1]

The physical installation made another connection between computing and music. It included The Hearth, an acoustic organ using the cooling fans associated with graphics processors. Air moved by a component of the computing machine became part of a musical instrument.[3]

In a November 2025 interview with Quincy Mackay for Reboot, Dryhurst explained the social arrangement of the room. Its square of benches drew on Sacred Harp singing: people face each other, and a caller in the centre both leads and receives the group’s response. The geometry gives the person prompting the music a particularly direct listening position.[3]

Herndon placed the project within a longer history of musical technologies, including notation, printing and the organ. It is a useful way to understand the exhibition’s mixture of old and new. A score already allows one person’s musical decisions to be interpreted by other people, across distance and time. Starmirror adds another stage of interpretation to that relationship.[3]

The voices have a future

An earlier project, The Call, had taken the pair and their collaborators to record 15 community choirs across the UK. For the Serpentine exhibition of 2024–25, they created a songbook of hymnal material and singing exercises, along with a recording protocol. Choristers also took part in an experiment about who should govern later uses of the data.[4]

Its white paper records a revealing exercise. Thirty-seven people cast 705 votes across a set of statements about possible uses. Their views differed, particularly on commercial reuse. The authors judged that an attribution, non-commercial, share-alike licence could fit the preferences expressed if the dataset were released.[5]

The white paper says the team ultimately selected Serpentine LLC, an existing trading entity, to act as the trusted intermediary. Choristers could shape future access through the licensing arrangements. The recording session therefore led to a further decision: who would receive requests to reuse the voices and apply the agreed terms.[5]

Those arrangements concern The Call. KW announced a public Starmirror dataset; its exhibition page supplies no download.[1][4]

Starmirror makes the beginning of an AI choir audible as a human activity: listening to a phrase, answering it together and making a recording. Before a model can sing back, somebody has to decide what to sing into it.[2]

Report an error or suggest a correction