ArrangementLabs was named the best project at the Demo Day of the fourth edition of Unicorn Hub Innovation Lab, where twenty-four teams presented their work. We demonstrated a tool for arranging music for different instrumental combinations: from uploading a piece and choosing the instrumentation to downloading the score. Producing the arrangement took about a minute and a half.
The first ideas for ArrangementLabs began to take shape in June 2025. We then devoted considerable attention to planning the concept and the system architecture in detail. We worked out how musicians would use the application, what material they could upload and in what form they would receive the arrangement. This informed the design of each stage of processing a piece and the way the individual parts of the system would work together.
The idea grew out of our own experience of arranging music for different instrumental combinations. The application offers two modes: transcription and creative arrangement.
Transcription mode transfers an existing score to the chosen instrumentation. It is intended, among other uses, for work with classical repertoire and early music, where preserving the original musical material as faithfully as possible matters. When there are fewer instruments than voices in the piece, some voices are combined.
Creative arrangement allows greater freedom to reshape the material. The model produces a new arrangement for the chosen instrumentation, using the description supplied by the user. It can alter the texture, voice leading and distribution of material among the instruments. This mode is for musicians looking for a different approach to a piece or a starting point for further arranging work.
In either mode, musicians can listen to the resulting arrangement, make corrections in the editor and export the score for further editing.
In our first trials, we turned a piano score into several instrumental parts. We listened to them and studied the full score to identify what needed improvement.
Alongside this, we worked on the quality of the arrangements. After the first round of additional model training on data relating to classical music, we began comparing arrangements produced by successive versions. Listeners assessing the results received several arrangements of the same piece, without being told which came from the newer version of the model. This allowed us to check whether our changes actually improved the arrangements.
Sometimes it was only when we listened without knowing the model version that we could no longer hear an improvement we had previously been certain of.
Such results made us reconsider our own ideas. Notation also took considerable time. The model could already distribute the material among instruments, but the resulting score was sometimes difficult to read and edit. We refined enharmonic spelling, beaming, the notation of pickups and the writing of two voices on a single staff. We corrected erroneous note values and worked on instrumental registers and the distribution of material among parts.
Every unclear bar meant work that a musician would have to do after us.
To assess the results, we also needed convenient access to the score and playback. We added the ability to listen to individual parts, mute instruments and compare an arrangement with the original. An editor followed, allowing users to change notes directly in the browser, hear the changes immediately and save a new version of the arrangement while retaining the previous one.
By mid-July, the catalogue included a dozen or so instrumental combinations: string ensembles, duos with piano, duos with accordion and solo instruments. Users could already go from uploading a piece through editing the arrangement to downloading the material in MusicXML, MIDI and PDF formats.
Demo Day gave us an opportunity to present the results of this work to a wider audience. Meanwhile, people from different parts of the world were joining the waiting list. We now want to learn about their repertoire, their needs and their responses to the arrangements.
The application remains in closed beta. We are sending invitations in batches and plan to release version 1.0 in autumn 2026. Ahead lies further work on the model, notation and arrangement editing, now with the involvement of musicians who will use ArrangementLabs in their own work.