MusicGen: Simple and Controllable Music Generation

Introducing MusicGen: Empowering Conditional Music Generation At Audiocraft, we have taken on the challenge of conditional music generation with our groundbreaking model, MusicGen. Unlike previous approaches, MusicGen is a single Language Model (LM) that operates seamlessly across multiple streams of compressed discrete music representation, represented as tokens. By leveraging an innovative single-stage transformer LM and efficient token interleaving patterns, we have eliminated the need for cascading multiple models, such as hierarchical or upsampling techniques. Through MusicGen, we showcase the ability to generate high-quality music samples, both mono and stereo, while leveraging textual descriptions or melodic features for enhanced control over the output. This approach revolutionizes the field of music generation, offering unprecedented flexibility and creative possibilities.