Not so long ago, the neural network of the domestic company Yandex recorded an album in the style of the Grazhdanskaya Oborona group. Many regarded this as an isolated occasion to draw attention to the company's developments. But, as it turned out, after recording in Russian, the programmers went even further and switched to foreign groups. For example, just the other day several tracks in the style of the American rock band Nirvana were recorded.
Since then, Yandex programmers Ivan Yamshchikov (who currently works at the Max Planck Institute for Scientific Mathematics in Leipzig) and Alexei Tikhonov have taught their neural network to write poetry in English, similar to the lyrics of the songs of Kurt Cobain, the leader of the world famous band. The first album of the neural network in the style of the "Civil Defense" group was recorded with the help of professional musicians and was called "Neural Defense". By analogy with the previous project, professional American musician Rob Carroll, a fan of the works of Kurt Cobain and his rock band, was hired to work on the recording of an album in the style of the Nirvana group. And the album itself was named "Neurona".
In order to teach the neural network to write poetry in English, Ivan Yamshchikov had to collect a huge archive of modern and classical foreign poetry with a total volume of about 200 megabytes. In the course of the work, it turned out that it is much more difficult to teach a neural network the Shakespearean language than Russian. The fact is that in English poetry there are much fewer clearly established rules and “normal” grammar does not always work there. Therefore, the team of experts had to delete some of the texts from the archive. After the scientists were satisfied with the result, they recorded the tracks on an old audio cassette with the Nirvana album. The final result was posted on the Web and even published a video for one of the songs on YouTube. As the developers assure, the algorithms of the neural network and the details of its development will soon be published in one of the scientific journals. Besides,the Russian edition of RIA Novosti managed to interview Ivan Yamshchikov, the full text of which, as well as a clip for one of the tracks authored by the neural network, can be found below.
Ivan, considering the difficulties that you faced while creating Neurons, have you tried to do something similar for the lyrics of the singer Linda, known for the unusual style of her poems?
A year ago, we trained a neural network to generate texts in the style of Yegor Letov. In principle, nothing prevents her from teaching her how to generate texts in Linda's style.
If neural networks can now even write poetry and music, is there something that they can never do in principle?
This is an interesting question for which there is no exact answer. The main thing to remember is that a neural network is not a character. It is simply a type of algorithm, a tool designed to solve a particular problem. How exactly to formulate the problem, conditions, restrictions, where to get data for training? For example, a person is working with all these questions.
What exactly made Neuron put you on an old cassette with Nirvana albums?
Promotional video:
We did the same with the Neural Defense tapes. It's kind of a tradition. There is something very familiar to the sound of a cassette player. It is not for nothing that a picture with a pencil, an old cassette and the caption “if you know how these objects are related to each other, you lived in the 90s” constantly circulates on the Web.
Given the emergence of neural networks capable of imitating the work of famous people, will there be a temptation in the future to introduce ownership and copyright not only for specific audio recordings, but also for the style of an artist, musician or composer?
Rather, this is a question for specialists in the field of law. In art, in general, there is still an element of exclusivity. The fact that neural networks are now able to imitate Van Gogh's style does not devalue his work. There is an element of mediated communication between the person who created this or that work, and the person who saw, heard, read.
VLADIMIR KUZNETSOV