How the Listen To Transformer neural network learned to create classical melodies on the piano

Recently, we have been releasing more and more interesting materials related to the work of neural networks. And no wonder: over the past 2-3 years, AI projects have reached a whole new level of development, forcing the whole world to follow breathlessly the results of IT labs dealing with the problems of developing fragments of artificial intelligence.

For example, we told you how a neural network implemented The main character Danila Bagrov from Alexei Balabanov's film "Brother" in the cult film "Home Alone. In a short video, Bodrov Jr. in the crowning knitted sweater entertains Kevin (Macaulay Culkin) on Christmas Eve. It turned out to be quite funny.

And here's a neural network inserted the face of a TV presenter Leonid Yakubovich into the body of a Chinese dancing teenager by picking up silly music. This may sound like child's play, but the fact remains: The video was produced by the robotwithout human involvement. And this is already interesting.

In today's article we will tell you how the artificial intelligence got to music as well. And this time the target of the neural network is not some pop or rap, but it has ambushed the sacred - classical music! And we will leave a link to it at the end of the article!

What is Listen to Transformer?

It should be noted that the developers at Magneta, an open-source research project investigating the role of machine learning as a tool in the creative process, have done their best. They have provided the world with radio neural networkthat infinitely generates melodies on the piano. Yes, you got that right - the artificial intelligence itself creates music, and does so countless times, and Not asking for a single penny!

There's a reason we call Listen to Transformer a radio-neural network: the interface really is a lot like a radio. By opening the page "Listen to the Transformer." on the site, we see the play bar, the song switch buttons, the pause button, the generated artwork (a simple picture of geometric colorful shapes is also generated by the neural network based on the notes itself), the numeric designation - the name of the currently playing track, and buttons with the ability to Save a favorite tune to your favorites and watch your favorite tracks.

What else is unique about it?

This neural network has several interesting features. First, it creates melodies instantly: It's easy to check this by simply switching to the next track in the player. It will play almost immediately, which tells us that this robot "thinks" very quickly. Second, the number of tunes generated no restrictions - you can create as many of them as you want, and each one will be completely unique! Third, the covers we mentioned are not formed in a chaotic order, but based on the length of the notes. This is how the developers explain this pattern:

The cover art for each song is generated algorithmically based on the notes in the song itself - while the notes are represented by random shapes, the degree of transparency is determined by the speed of playback, and the size of the shapes depends on the duration of each note used in the tune.

In addition, the developers of this neural network justify themselves: it is possible that the tunes you hear are not as good as the ones that people create, but it is interesting that it generates artificial intelligence. And interested musicians can absolutely free take any samples (musical segments) they like and insert them into their works, modify them and do with them whatever they want. After all, the rights to the music created by the cyborg are not protected - The authors of "Transformer" mention it.

The Fuzz Music team has listened to about 30 tunes created by this neural network, and can confirm that at best half of them have some more or less acceptable melody. Unfortunately, the rest of this music is not of much interest, except for programmers and developers in the field of artificial intelligence.

By the way, the Magneta developers have their own own blogwhere they share the results of their research on machine learning tools. It's all in English, but if anyone is interested, here it is. Let's follow the developments and find out how modern neuro-technologies will influence music and the music industry! Maybe in the future similar projects will be able to create harmonious melodies and make us happy!