A year ago, “artificial intelligence” from “Yandex” composed poems in the style of Yegor Letov
, which they then put on music in the style of the “Civil Defense” group. This was the first musical experience of Avtopoet, developed by the company ten years ago. Since 2006, he tried himself in writing poems for Pushkin, poems based on search queries, he was used along with the voice generation system to recite poetry in the author’s performance.
Now program developers Alexey Tikhonov
and Ivan Yamshchikov
(currently working at the Max Planck Institute for Scientific Mathematics Institute in Leipzig) decided to repeat the experience and do something similar, but in English. As a model to follow, “Autopoetu” was given the cult group Nirvana. Especially since Kurt Cobain would have turned 50 this year.
Ivan Yamshchikov writes
that they collected about 200 megabytes of classical and modern poetry all over the Internet to train a neural network. But suddenly it turned out that it was impossible to teach a neural network using these raw data: “Obviously, many people write poems in English, but about 50% of them do not know the basic rules of English grammar. The most basic things, ”laments Ivan.
The collected data had to be carefully filtered, which took a long time. The data set was reduced approximately by half, that is, to approximately 100 MB of texts. After filtering, the quality of the data was sufficient to train the language model, and the neural network was then specially trained to follow Kurt Cobain.
At the same time, the authors published a preprint of scientific work
on the generation of music by the neural network. Most likely, the algorithms of the neural network and development details will soon be published in one of the scientific journals.
True, for the four songs of this album in the style of "Nirvana", Yamschikov himself wrote the backing track, not the neural network, but the program composed the verses. Of the entire volume of poems composed by her, the best ones were filtered out and selected. Then the authors began to look for a vocalist who would perform these works. Four performers were listened to, but only the fifth satisfied the rigorous requirements - Rob Carroll
, he recorded vocals for all four songs.
As in the case of “Civil Defense”, for greater realism, the authors transferred the result to an old film cassette, and then digitized the sound from it with a characteristic hiss, crackle and distortion. They say that we all listened to “Nirvana” on old tape recorders, so this sound will seem more familiar and familiar.
The process of recording on film is shown in the video
. The tracks were simply launched on the iPad in the meter from the equipment that recorded the sound on the tape. For the sake of the symbolic meaning, the songs were recorded not just on some random tapes, but on tapes with Nirvana albums, on top of old songs.
So, you can yourself evaluate the result. The Model
album from the Neurona
team has been published on all popular music platforms.AppleMusic:https://itunes.apple.com/album/id1248121490Google Music:https://play.google.com/music/preview/Bbeknqo5djw3kysh7ildla6zgneSpotify:https://open.spotify.com/artist/6Y3iq1wJEPEj6e6P0NNL4ZYandex.Music:https://music.yandex.ru/artist/5153458
In one of the songs from the album In the back of your glass they
even made a video clip
Earlier on GT and Habré it was reported about scientific works in which neural networks successfully transfer the style from one photo to another
, as well as the style of the artist from one picture to another
As we see, neural networks can transfer both the style of poetry and the style of music. In an interview with
RIA Novosti, Ivan Yamshchikov was asked a question, 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 to specific audio recordings, but also to the style of an artist, musician or composer? The author of the autopoiet program does not see such a threat. Quite the contrary, he believes that AI can be a co-author and help a person: “In art as a whole, there is still an element of exclusivity. The fact that neural networks are now able to imitate the style of Van Gogh 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 it, heard it, read it. This, I think, should remove the copyright issues. Much more interesting is that machine intelligence can become a co-author, another tool that is simply available to the artist and allows you to do more interesting things. ”