Jukebox

Includes an exploration tool.

It is customizable with user-specified genre, artist, and lyrics.

Capable of generating music unrelated to its training data.

Has the ability to condition on brief audio snippets.

Models music directly as raw audio.

More versatile and expressive than symbolic music tools.

Incorporates diversity and long-range structures.

Can compress raw audio.

Simulates music and melody.

Replicates genre and artist style.

Creates original music.

Generates basic singing.

Offers multi-genre support.

Uses autoencoder for audio compression.

Employs VQ-VAE for audio compression.

Applies Sparse Transformers for autoregressive modeling.

Balances local musical structures in the music.

Outputs high-quality raw audio.

Expands the possibilities for generative models.

Can produce coherent songs of significant length.

Adapts to various music and singing styles.

Manages the complexities of raw audio sequences.

Capable of creating completely unique music.

Captures high-level musical semantics.

Captures timbre, melodies, and dynamics.

Produces a diverse range of music output.

Models raw audio directly.

Its autoencoder compresses raw audio sequences.

Model weights and code are released.

Learned to group similar artists and genres.

Can be conditioned on artist and genre.

Includes lyrics conditioning.

Aligns lyrics characters with song duration.

Artist and genre conditioning.

The EncoderDecoder attention layer learns LyricsMusic Alignment.

Matches audio segments to corresponding lyrics.

High musical quality when compared to comparable tools.

Enhanced sound quality with scaled VQ-VAE.

Creates long, coherent songs.

The model learns to integrate additional conditioning information.

Experiences loss of audio details.

Generates noticeable noise.

Generates songs at a slow pace.

Lacks repeated chorus structures.

Less useful for musicians.