Facebook Research introduces MusicGen, a single Language Model for conditional music generation. Unlike previous systems that require cascading several models, MusicGen operates over several streams of compressed discrete music representation and uses a single-stage transformer LM with efficient token interleaving patterns. It can generate high-quality music samples while being conditioned on textual description or melodic features. The proposed approach is superior to the evaluated baselines on a standard text-to-music benchmark. Music samples are available on the supplemental materials, and the code and models are available on the team’s GitHub repository.
https://ai.honu.io/papers/musicgen/