Source: Max Pixel
Written by | Ahmed Elgammal (Professor, Director, Art and AI Lab, Rutgers University)
Compile the | Måka
1
In 1827, when Ludwig van Beethoven died, one of his masterpieces, the Ninth Symphony, was published less than three years ago.
Around 1817, the Royal Philharmonic Society of London commissioned Beethoven to compose the Ninth Symphony and the Tenth Symphony. By 1824, he had completed the Ninth Symphony, which ended with the familiar Ode to Joy.
Subsequently, Beethoven began to compose the Tenth Symphony. But as his health continued to deteriorate, the work failed to make much progress, and the great musician ended up with only notes, fragments, and sporadic ideas he had scribbled down.
Some of Beethoven's notes on the Tenth Symphony. | Image credit: Beethoven House Museum, CC BY-SA
Beethoven's fans and music theorists have always regretted this. In the past, there have been attempts to reconstruct some parts of the Tenth Symphony. One of the most famous earlier attempts was in 1988, when music theorist Barry Cooper ventured to complete the first and second movements. He used notes left by Beethoven to weave 250 bars of music, which in his view was faithful to the first movement of Beethoven's ideas.
But because Beethoven's drafts were so limited, even symphony experts had a hard time reconstructing more parts.
Now, with a team of music historians, music theorists, composers, and computer scientists working together, the void left by Beethoven is about to be filled, and his unfinished Tenth Symphony is about to be completed. After more than two years of hard work, the full recording of Beethoven's Tenth Symphony will be released on October 9, 2021.
2
In early 2019, to mark the 250th anniversary of Beethoven's birth, Dr Matthias Röder of the Karajan Institute in Austria planned to assemble an interdisciplinary team to try to complete Beethoven's Tenth Symphony.
This team is mainly divided into two parts. The Artificial Intelligence (AI) team is largely led by Professor Ahmed Elgammal, director of Rutgers' Art and AI Laboratory, whose mission with a group of computer scientists is to teach AI Beethoven the creative process.
The music team was joined by a number of famous musicians, including the Austrian composer Walter Werzowa. One of Wozova's most widely circulated creations is Intel's iconic advertising track. Computational music expert Mark Gotham led the work of trancribing Beethoven's work. The team also included music theorist Robert Levin, who was also a skilled pianist who had previously completed many unfinished works by Mozart and Bach in the 18th century.
In simple terms, the whole team needed to use notes and finished works of All Beethoven, plus a draft of the Tenth Symphony, so that the machine could learn to continue to compose.
3
It's a huge challenge because there's no off-the-shelf machine that can just type in those drafts, press a button, and spit out an entire symphony. At the time, most of the available AI couldn't continue to expand unfinished music beyond a few seconds.
The team needed to find a way to teach the machine Beethoven's creative process to push the limits of what creative AI could do. With the help of training on a large number of Beethoven's works, the team allowed them to learn how Beethoven started with a few bars of music and painstakingly composed them all the way into exciting symphonies, quartets and sonatas.
The music team cracked and transcribed a draft of the Tenth Symphony and tried to understand Beethoven's own creative intent. Using Beethoven's completed symphony as a template, they tried to piece together where fragments of the draft should be placed, such as which movement they belonged to, or which part of the movement they should be placed in.
Symphonies usually consist of 4 movements, and the team must make decisions based on music theory, such as determining whether a draft represents the beginning of a harmonic (i.e., a very vivid part of the symphony), or they may want to confirm that a certain line of music is the basis of a fugue. A fugue is a melody composed of intertwined parts, all of which echo a central theme.
AI teams, on the other hand, struggle to complete a series of challenging tasks. For scientists, this challenge is actually daunting. Because to be successful, AI needs to accomplish tasks they've never tried before.
First and foremost, the team needed to figure out how to develop a longer, more complex musical structure from a short clip, or even just the purpose of a piece, as Beethoven did when he was creating it. For example, the machine had to learn how Beethoven constructed the Fifth Symphony with a basic four-note note tone.
Second, the continuation of the passage also needs to follow the laws of certain musical forms, whether it is a harmonic, trio, or fugue, and AI also needs to learn the process by which Beethoven developed these forms.
The team must also teach the AI how to master melodic lines and harmonize them. In addition to this, the AI also has to learn how to connect the two pieces of music and create the ending part.
Finally, once a complete piece is available, the AI also tries to orchestrate it, which involves assigning a variety of instruments to different parts.
All in all, throughout the process, the machine must learn to complete various tasks like Beethoven.
4
In November 2019, the team gathered at the Beethoven House in Bonn, a meeting that was a litmus test to determine whether AI could complete the project.
They printed out the sheet music created by the AI based on the draft left by Beethoven and invited a pianist to play in a small concert hall in the House Museum. The audience included journalists, music theorists and experts on Beethoven.
The team challenged the audience to determine where the passages created by Beethoven ended and where the additional creations of the AI began. In fact, the audience was unable to distinguish it.
A few days later, at a press conference, part of the score was again played in public. Only those who are very familiar with the draft of the Tenth Symphony can determine which parts were generated by AI.
The success of these tests shows that the team's research direction is correct. Subsequently, they continued their efforts to develop a complete score for the Tenth Symphony.
Beethoven's unparalleled talent challenges all researchers at all times, and it requires the team to do better. As the entire project evolves, SO does AI. Over the next 18 months, the team built and organized two complete movements, each lasting more than 20 minutes.
The team is also aware that there may be some different voices about this study. Some would argue that art should be off-limits to AI, and there is no need for AI to try to replicate the human creative process. But when it comes to art, computer scientist Professor El Jamal believes that AI is not a substitute, but a tool that allows artists to express themselves in new ways.
This project would not have been possible without the expertise of human historians and musicians. At one point, a music expert on the team said that AI reminded him of a student who was eager to learn music, who practiced, studied, and got better and better every day.
Now, the student has taken beethoven's baton and is ready to play his Tenth Symphony to the world.
Reference Sources
https://theconversation.com/how-a-team-of-musicologists-and-computer-scientists-completed-beethovens-unfinished-10th-symphony-168160
This article is reproduced with permission from the WeChat public account "Principle".