AudioCraft is Meta AI’s PyTorch library for deep learning research on audio generation, best known for MusicGen, its controllable text-to-music model, and AudioGen for text-to-sound. It also includes EnCodec, the high-fidelity neural audio codec that underpins the generative models, plus newer additions such as MAGNeT, AudioSeal watermarking and the JASCO music model. With about 23,700 stars on GitHub, it is the reference open-source toolkit for generative audio research.
AudioCraft provides both inference and training code, which sets it apart from demo-only releases. MusicGen generates music from text descriptions, optionally conditioned on a melody, using a single language model over EnCodec tokens; AudioGen applies the same approach to sound effects and environmental audio. Around these sit supporting models: Multi Band Diffusion as an alternative EnCodec decoder, MusicGen Style for style-conditioned generation, MAGNeT for faster non-autoregressive generation, AudioSeal for proactive audio watermarking, and JASCO for generation conditioned on chords, melody and drums. Models are distributed through Hugging Face, documentation and model cards are included, and the codebase is designed so researchers can train, fine-tune and extend the models rather than only run them.
AudioCraft is free and open source: the code is released under the MIT license. Note that the model weights use the CC-BY-NC 4.0 license, so commercial use of the weights themselves is not permitted under that license.
AudioCraft is for audio researchers, ML engineers and creative developers who want to generate music and sound with open models, study how neural audio codecs and generative models work, or build prototypes on top of Meta’s audio research.