Research laboratory OpenAI from San Francisco (USA), which specializiruetsya on the development of artificial intelligence, presented a new successful project. He got the name “Jukebox” (jukebox) and is intended for meticulous study and subsequent playback of votes and manner of performance of real singers. Training began with the works of Sinatra and Elvis and Jukebox showed great results.
Jukebox works as a synthetic Duo composer and singer-songwriter that strikingly distinguishes it from live musicians. On the one hand, he correctly combines notes and sounds, on the other – still makes stupid mistakes, “wrong” in how to play individual pieces. Because AI he writes music from scratch and he is singing words as the input data it is given only the text that still need to analyse and understand how to sing.
Texts created another artificial intelligence, special language model, edited by living members of OpenAI. They correspond roughly to the creativity of the selected artists – Jukebox was trained on 1.2 million individual tracks. The feature of this AI is that when listening to a song he splits it into three levels. The first is the maximum compressed bit rate, just the melody and arrangement of words. The next two levels add detail and specific features of the particular musicians and singers.
When the Jukebox starts to simulate the song, he goes the same way. First builds a “skeleton” and then saturates its parts to the desired level. As it happens, only solves the AI, which explains the still high probability of errors. But where the Jukebox sings well, even the most experienced listener are unlikely to suspect a trick, unless he knows the work of the same Sinatra, and firmly believe that such a song he never sang. However, you can do experiments in a series of “what if an Opera singer of the last century would charge to sing the latest pop hit?”. Examples of the art of artificial intelligence can be streamed on the project website.
Source — OpenAI