Jukebox

OpenAI · 2020

Synthesizes raw audio songs with singing vocals conditioned on genre, style, and lyrics.

Application types: Text-to-musicLyrics-to-song
Architecture:Autoregressive transformer

Essential categories

CategoryStatusNotes
Source code✔︎ OpenCode is available at GitHub repository. The codebase is not actively maintained anymore, but it is still available for use. It includes the model architecture, training and evaluation code, and utilities for generating music samples.
Training data~ PartialTraining data is partially described in the preprint article. However, relevant information is missing, such as specific sources, and data itself cannot be accessed. Dataset is not publicly available.
Model weights✔︎ OpenWeights are available.
Code documentation✔︎ OpenCodebase is documented.
Training procedure✔︎ OpenInformation available between article and GitHub repository.
Evaluation procedure~ PartialPart of the evaluation procedure is documented in the research paper. However, key information is missing, such as the evaluation dataset, objective evaluation metrics and details on the manual evaluation.
Research paper~ PartialOnly preprint of the research paper is available.
Licensing~ PartialCodebase is licensed under a non-commercial use license by OpenAI.

Desirable categories

CategoryStatusNotes
Model card∅ Not includedNot available.
Datasheet∅ Not includedNot available.
Package∅ Not includedNot available. Codebase is officially not mantained.
User-oriented application∅ Not includedOnly a colab notebook is available.
Supplementary material page⭐ IncludedA demo/summary page is available. However, most sonifications of the examples generations are not available anymore. There exist another site with generation examples, but most of them are not available https://jukebox.openai.com/.

Raw YAML file with complete evaluation.