Hyppää pääsisältöön


The AI judge of the Sibelius Violin Competition did not follow the human jury’s lead – it made its own decisions

Vuodelta 2022
Armenialainen viulisti Diana Adamyan soittaa orkesterin edessä katse suunnattuna yläviistoon ja jousikäsi kohotettuna.
Kuvateksti Diana Adamyan, finalist in the Sibelius Violin Competition
Kuva: Minna Hatinen

An AI judge was trialled at the final of the Sibelius Violin Competition. Diana Adamyan of Armenia was given top marks by the AI. The actual winner of the competition, Inmo Yang of South Korea, was given a second-place.

Can artificial intelligence learn to evaluate a violin performance? This interesting question was put to the test in the finals of the international Jean Sibelius Violin Competition, where each of the six finalists played Sibelius’ Violin Concerto as a compulsory piece.

Four parts of the Sibelius Concerto were taught to the algorithm during the spring. The features to be taught included clarity, sound quality, cooperation with the orchestra and handling of the rhythm. Previous winning performances from 1965–2015 were used as the teaching material.

1st prize Adamyan, 2nd prize Yang, 3rd prize Moroz

The report produced by the AI indicated that the winner of the playful trial competition was Diana Adamyan of Armenia. The AI judge awarded second prize to the actual winner of the competition, Inmo Yang of South Korea, and third prize to one of the Ukrainian finalists, Georgii Moroz.

Diana Adamyan: Sibelius Violin Concerto, 3rd movement - Toista Yle Areenassa
Inmo Yang: Sibelius Violin Concerto, 3rd movement - Toista Yle Areenassa
Georgii Moroz: Sibelius Violin Concerto, 1st movement - Toista Yle Areenassa

This year, the Sibelius Violin Competition was of an exceptionally high standard, certainly putting the eight-member jury to the test. The task of the jury was to evaluate the entire competition performance, while the AI focused only on part of one competition performance.

The AI report also shows that the scores between the six finalists are extremely close. There are differences, but they are minute.

Äänen spektrogrammi
Kuvateksti The beginning of the 3rd movement of Sibelius’ Violin Concerto - (top) Diana Adamyan and (bottom) Inmo Yang

The AI test design was challenging

Recordings of previous winning performances of the Violin Competition were used as the teaching materials for the artificial intelligence. Sakari Oramo, chairman of the jury, listened to and ranked these recordings.

Because all of the previous winners were good, the algorithm learned how the Concerto should be played. With regard to learning, the algorithm would also have to be taught the performances in which the violinists did not perform as well.

Another challenge to algorithm learning is that each of the previous winning performances is a product of its own time. For example, in 1965, recordings were made with completely different technology and different musical ideals were held.

Even though background noise in the old recordings was removed and the recordings were improved, they still cannot be fully compared with today's recordings due to their qualitative differences.

The third factor hampering learning was that, if musical approaches and interpretations were being critiqued, these would also have to be included in the material being taught. When the neural network analyses an individual performance, it looks for similarities between it and the material taught.

Äänen spektrogrammi
Kuvateksti Spectrogram: (top) 1985 winner Leonidas Kavakos and (bottom) 2022 winner Inmo Yang, 1st movement of the Sibelius Violin Concerto

The algorithm learns in the same way as a human

The algorithm learning process consists of phases. Initially, it looks for similarities in the teaching materials in order to confirm that the input music recordings belong together - in other words, that it is the same piece of music, the Sibelius Violin Concerto.

Teaching an algorithm is done in much the same way as teaching a human. It is taught by hundreds and hundreds of repetitions.

Self-learning algorithms are characterised by the way they develop with the help of reliable observations. As the learning process progresses, the ability of the algorithm to determine whether or not the performance matches the reference data increases, and recognition of the performance gradually becomes secondary.

The algorithm begins to look for minor differences in music performances outside the basic observations.

How did the playful experiment do?

The AI results obtained raises the question as to whether AI could replace a human judge in the Violin Competition? There is no one right answer, but, for the time being, human judges can breathe a sigh of relief.

Their work will continue in the years to come, as artificial intelligence learning will take time.

What is positive, however, is that artificial intelligence does indeed learn. The learning curve is constantly improving.

The AI judge was part of content produced by Yle for the Sibelius Violin Competition. AISpotter was responsible for the technical execution of the experiment.

Sisältö muilla kielillä