Music, the last frontier of Artificial Intelligence

You may already have heard of Eternity, the first K-Pop group made up entirely of members created by artificial intelligence. With its epicentre in South Korea, this musical trend has managed to become global and make millions of fans thanks to the strength of groups like BTS and Blackpink, and songs composed with the help of algorithms based on the public tastes that create success after success. So far, so good, but listen to this: None of the eleven members of Eternity exists. (But neither does Hatsune Miku who triumphed at the Wizink Center in 2020. People have even married her. And been divorced from her.)

If you did not already know, you may never have realised. On their social networks they look real, but not so much in their video clips. The voices (and bodies, apparently) are real, but the people dancing are avatars created with deepfake technology. 

The truth is that there is no shortage of advantages from the business point of view: Virtual artists do not have the physical or mental limitations of humans. The General Manager of Pulse9, which is behind Eternity said that this project is creating "a new genre" but, to what extent is this accurate?

Since the best idea is to go to the source, we decided to ask AI directly. This is the reply we got from ChatGPT when we asked if it could compose a song.

Yes, it is possible for AI to compose music. There are a number of different approaches to achieving this, these include supervised and unsupervised machine learning, and the generation of algorithms. A common approach is to train a machine learning model with a large number of music examples, and then use it to generate new music that resembles the examples used for training.

There are some existing applications and projects that use AI to generate music, but they are still in the experimental phase and the result is not always very convincing. But as this technology and machine learning progress, it is expected that in the future it will be possible to compose music with AI.

That is, Artificial Intelligence relies on examples of songs to compose other songs that resemble the originals. It cannot be done from scratch without references. If the truth be told, neither can people.

Things get complicated when we asked it if it could create a new style of music. Here is the answer: It is possible for AI to create a new musical style, but this would depend on the approach and complexity of the model used. AI could generate music that combines elements of various existing styles and create something that is unique and unprecedented. However, for AI to create a new musical style autonomously, it would require a great knowledge of music theory, aesthetics and conventions with regard to different musical styles, as well as the ability to generate melodies and arrangements that fit that style.

It is also important to mention that music is not only a technical issue, but there is also a cultural, social and emotional dimension, so generating a new musical style is not only a matter of generating random sounds but of being able to generate something that connects with people.

In this answer, something comes into play that is not within the reach of AI (for now): The emotional dimension. Music, and music lovers know all about this, is above all emotion.

Everything became clearer with the last question we asked: Do you like music? As AI, I have no personal preferences or tastes, so I cannot tell you whether or not I like music. However, I am designed to provide information and help when it comes to music or any other topic. If you have any specific questions about music, I will be delighted to help you.

To summarise, Artificial Intelligence will compose a song based on a reference, but it is unable to do so to express feelings or aesthetically. It simply cannot, because it cannot know whether or not it likes what it is doing (again, for now).

So what role does AI play in music? Well, basically it can do the dirty work. A kind of calculator that provides you with the square root you need to reach an objective: This could involve studying thousands of greatest hits and finding the keys that enabled them to reach the top of the charts, generating neutral music free of rights to use on social networks as proposed by the startup Jukedeck, or simply to be surprised with what it is capable of doing. Take a quick look at Jukebox, a cousin of ChatGPT that generates musical pieces in the style of the artist chosen (voice included) as a result of analysing a database with more than a million songs.

Here are some examples

If you would like to experiment with this, here are some interesting things to look at: For example, the Moises AI application, with which you can do things like removing the voice from your favourite songs to create a personalised karaoke. And while you are at it, you could try Riffusion, which generates a small melody from a phrase you enter. Watch out, it is addictive. These are just some examples for simple users, but there are also tools for creators, including Magenta, by Google, and Sony with Flow Machines.

So, endless possibilities for all needs. Internet in its purest form. The sky is the limit!