Jukebox is an OpenAI research project that generates music — including singing — in a wide range of genres and styles, entirely from scratch using AI. It was published in 2020 as an open-source demonstration but was never turned into a product. It is likely no longer actively maintained.

Jukebox was trained on 1.2 million songs and uses a technique called a VQ-VAE (vector-quantized variational autoencoder) to compress audio into a format that an AI model can learn from and reconstruct. It can imitate the styles of artists across rock, hip-hop, jazz, pop, and classical music. A significant practical limitation is that it takes approximately 9 hours of computing time to generate one minute of audio. OpenAI released the model weights and source code publicly under an open-source license, but acknowledged a significant gap between Jukebox's output and music made by human artists.