jukebox

Definition

An OpenAI model that generates raw-audio music including vocals. It compresses audio with a multi-level VQ-VAE into coarse-to-fine discrete codes, models each level with autoregressive Transformers, and upsamples back to a waveform; genre, artist, and lyrics can be supplied as conditioning. The results were striking for the time but slow to sample and audibly rough — costs that pushed later systems toward neural audio codecs and latent diffusion instead of direct autoregressive waveform modelling.