Jukebox: A Generative Model for Music

TL;DR

Jukebox is a deep generative model that produces high-quality, diverse music with singing directly from raw audio, using multi-scale compression and autoregressive modeling, allowing style and lyric control.

Contribution

The paper introduces Jukebox, a novel model combining multi-scale VQ-VAE and Transformers to generate coherent, multi-minute music with singing, conditioned on style and lyrics.

Findings

Generates high-fidelity, diverse songs up to several minutes long

Achieves controllable singing with style and lyric conditioning

Provides publicly available samples, code, and model weights

Abstract

We introduce Jukebox, a model that generates music with singing in the raw audio domain. We tackle the long context of raw audio using a multi-scale VQ-VAE to compress it to discrete codes, and modeling those using autoregressive Transformers. We show that the combined model at scale can generate high-fidelity and diverse songs with coherence up to multiple minutes. We can condition on artist and genre to steer the musical and vocal style, and on unaligned lyrics to make the singing more controllable. We are releasing thousands of non cherry-picked samples at https://jukebox.openai.com, along with model weights and code at https://github.com/openai/jukebox