OpenAI released Jukebox on 30 April 2020, a neural network designed to generate music with singing in raw audio form [source]. The model produces complete songs including rudimentary vocals, accompaniment, and genre-specific characteristics across styles from rock to hip-hop.

Jukebox operates by compressing audio into a discrete space using VQ-VAE, then training transformers to generate music in that compressed representation. The system can generate minute-long samples at 44.1 kHz, though OpenAI acknowledged significant computational requirements and generation times measured in hours per sample.

The model accepts conditioning inputs including genre, artist, and lyrics. OpenAI reported that Jukebox could capture artist styles and generate novel compositions in those styles, though the company noted limitations in musical coherence over longer timescales. Generated vocals remain difficult to understand, with lyrics often unintelligible despite conditioning.

OpenAI released sample outputs demonstrating the model's capabilities across multiple genres and artist styles. The company also published code and model weights for research purposes, though noted that generation requires substantial GPU resources.

The release represented an early attempt at end-to-end neural audio generation at scale. OpenAI framed Jukebox as a research milestone rather than a production system, citing ongoing challenges in audio quality, coherence, and computational efficiency. The company indicated that future work would focus on improving sample quality and reducing generation time.

No specific failure modes or unexpected behaviours were documented in the initial release announcement. OpenAI presented the system's limitations as known constraints of the approach rather than operational issues requiring remediation.