AI that automatically generates songs with lyrics

OpenAI, a non-profit organization that studies artificial intelligence, has unveiled Jukebox, an AI that uses neural networks to create songs containing specified genres, artists, and lyrics.

The total number of published samples was 7,131 songs. According to OpenAI, there have been attempts to create music with machine learning so far, but the problem was that the amount of data contained in the music was too much to interpret. To solve this problem, the jukebox perceptually discards the intermittent information from music and compresses the raw data with a convolutional neural network. Music is created from this compressed data and then upsampled to restore the original sound quality.

To train the jukebox, we use 1.2 million songs as a dataset. The dataset music was divided into lyrics, artist, and album, and metadata such as genre and keywords were linked. Through various information such as metadata, the jukebox classifies artists by genre. Through diagrams, artists in close proximity are classified as having close relationships. The classification table on the official page of the jukebox tells us this relationship.

The OpenAI research team said that although jukeboxes have advanced one step in their abilities such as music quality and consistency, audio sample length, artist, genre, and lyrics, there is a big gap with human-made songs. It includes chord patterns and solo parts, but acknowledges that there are limitations of jukeboxes by saying that there is no extremely general musical structure such as chorus repetition.

The jukebox source code is also available on GitHub. Related information can be found here .