AudioCraft / MusicGen

MusicGen is the text-to-music model at the center of AudioCraft, Meta's PyTorch library for generative audio research. Given a text description, and optionally a melody to condition on, it generates music using a single-stage autoregressive transformer over EnCodec audio tokens. The AudioCraft repository packages it alongside AudioGen for text-to-sound generation and the EnCodec neural codec, with both inference and training code so researchers can fine-tune or retrain models on their own data. Pretrained MusicGen checkpoints come in several sizes, and running the larger ones comfortably calls for a GPU with several gigabytes of VRAM. The licensing is split: the AudioCraft code is MIT licensed, but the released model weights are distributed under CC-BY-NC 4.0, which excludes commercial use. Musicians, researchers, and hobbyists use it for prompted composition sketches and melody-guided variations, and the models are also accessible through the Hugging Face Transformers library.