OpenAI presented, last Thursday (30), a new model of artificial intelligence capable of creating complete songs, including the instrumental part, lyrics and vocals. Entitled Jukebox, the open source system showed impressive results when compared to others of the type already released.

According to the organization, a massive set of musical data was used to train the Jukebox, covering almost countless genres. The goal was to make the neural system create completely new songs, from known works.

Based on the songs used to train her, artificial intelligence can imitate style and genre, among other characteristics. It is also able to recreate the vocalist’s lyrical and singing style, presenting a very different result from another tool launched by OpenAI, called MuseNet, which produced only basic MIDI tracks.

Despite the advances obtained, especially in relation to the structuring of the songs and the duration, the novelty still has some limitations, mainly in the vocal part. Another detail is that the system takes about nine hours to render a minute of audio, making it impossible to use it in interactive applications.

Listen to the songs created by AI
The data set used to train the artificial intelligence Jukebox covered 1.2 million songs, of which 600 thousand were in English. The researchers also used the lyrics of the songs and metadata taken from LyricWiki, including information such as artist or band name, genre, release year and keywords associated with each track.

The entity made the songs created by the AI ​​available on the website There are more than 7 thousand songs created by the system and to hear them just click on the desired track, inside the page, according to the name of the artist or genre.

Among the artists and bands that served as inspiration for the tool, are names like Frank Sinatra, Céline Dion, Miley Cyrus, Ed Sheeran, Rihanna, Ramones, NOFX, Eminem, Bob Marley, Jethro Tull, Katy Perry and Kanye West.


Please enter your comment!
Please enter your name here