Jukebox

Jukebox, released by OpenAI in April 2020, was the first system to generate music with singing directly in the raw audio domain. It compresses 44.1kHz audio into discrete codes with a hierarchical VQ-VAE at three temporal resolutions, then models those codes with autoregressive sparse transformers, producing minute-long songs with long-range coherence.

Generation can be conditioned on artist, genre, and unaligned lyrics to steer musical and vocal style. OpenAI released thousands of non-cherry-picked samples along with model weights and code, making it a landmark of open audio research. It is enormously compute-hungry by modern standards and has been surpassed in quality by diffusion and codec-LM systems, but as a historical entry it proved raw-waveform song generation was possible at all.