Jukebox

What is Jukebox?

Jukebox is a neural network by OpenAI that generates music, including rudimentary singing, as raw audio. It accepts genre, artist, and lyrics as inputs to produce new samples from scratch. The model weights and code are released for exploration, though rendering takes approximately 9 hours per minute of audio.

What are the features of Jukebox?

• Raw audio generation

  Produces music directly as 44.1 kHz raw audio rather than symbolic formats like MIDI or piano rolls.
• Style conditioning

  Accepts artist and genre inputs to steer the generated output toward specific musical styles.
• Lyrics conditioning

  Takes text lyrics as input and aligns them to the music using an encoder-decoder attention mechanism.
• Audio completion

  Continues a track when provided with 12 seconds of conditioning audio in a specified style.
• Hierarchical compression

  Uses a three-level VQ-VAE autoencoder to compress audio by 8x, 32x, and 128x before generation.
• Sample explorer

  A released tool to browse uncurated generated audio samples.

What are the use cases of Jukebox?

• Song continuation

  Provide 12 seconds of existing audio to generate the rest of a track in a chosen style.
• Unseen lyric performance

  Input original lyrics co-written by researchers and a language model to hear them sung in a selected genre.

Jukebox Pricing Information:

• Open-source release

  Model weights and code are released publicly; no pricing is stated.

Jukebox Contact Email:

jukebox@openai.com