In 2020, Ludwig van Beethoven would have turned 250 years old. On September 2 and 3, 2021, in Lausanne and then in Geneva, the Swiss orchestra Nexus celebrated this anniversary, one year late due to a pandemic, with two concerts. In a program with Brahms and Rachmaninoff, he performed an unpublished work by the German composer. Honestly, a work that he did not actually write: his 10th symphony, of which only preparatory fragments remain, some of which were attributed by experts to a future symphony without Beethoven having explicitly referred to it.
Thus, a reconstruction of the first movement was carried out and it was played in 1988. But the Nexus orchestra started from another base: artificial intelligence tools capable of automatically generating scores from a training database.
Requested by the director of the Nexus orchestra Guillaume Berney, it was researcher Florian Colombo, from the Federal Polytechnic School of Lausanne (EPFL), who led this project. Specialist in artificial neural networks and in “deep learning”, cellist, works in compositional aid technologies. His work led to the BachProp Project, which involves producing scores with a particular style of music and musician.
A concert on October 9, 2021 in Bonn
At the same time, Deutsche Telekom is also aiming for Beethoven’s 10th Symphony to be written by artificial intelligence. A concert is scheduled in Bonn on October 9, 2021. And in the same vein, the Chinese telecommunications equipment manufacturer Huawei had “completed” Schubert’s Eighth Symphony in 2019. But in both cases, the algorithms create music, which then it is modified, orchestrated and partitioned by real musicians. “There my idea is to see how far we can go in automation”, summarizes Florian Colombo. His technologies, around which start-up AdaMu built, provide written scores for every instrument in the work.
The researcher worked from a learning database made up of Beethoven’s 16 string quartets. He had already used them for a previous project and took them up again, believing that they well represented the master’s style of composition. “It also allowed me to go faster, admits Florian Colombo. If I had had more time, I would have used the sheet music for the symphonies.”
However, this database contains information related to all grades. For each instrument, three main criteria are analyzed: melody (that is, what the first violin plays), rhythm, and harmony. From which the neural network is able, based on pieces of existing scores, to “predict” what would be the continuation, respecting the style of a specific composer. The algorithm, baptized as part of this BeethovANN project, was subjected to fragments written by Beethoven for his 10th symphony and generated musical ideas with its transcriptions. These fragments were retrieved from the IMSLP online library.
A choice between several “probable sequences” of the score.
However, the role of humans is not reduced to nothing: the algorithm can produce several predictions to choose from. This choice determines the next predictions, the “probable continuation” of the score. “Initially, all the criteria have the same weight, explains the researcher. But you can force an instrument to play this or that fragment, and the others are built around it. Depending on the decisions you make, the odds change. It all depends on what you start with. “In this case, Florian Colombo started by generating the melody and then came the bass. But the melody could have been different if he had played the bass first. Either way, the neural network does not follow any rules for musical composition is about training from data.
This project is, in fact, just one step. Florian Colombo intends to continue developing a true compositional tool, precisely including elements of music theory. With this ambition, in short, that the musician could interact a little more with technology.