AudioCraft

AudioCraft is Meta's PyTorch library for research on generative audio. It bundles inference and training code for a family of models: MusicGen for controllable text-to-music, AudioGen for text-to-sound generation, the EnCodec neural audio codec for high-fidelity compression, MAGNeT for non-autoregressive music and sound generation, Multi Band Diffusion decoding, the AudioSeal watermarking tool, and JASCO, which conditions music generation on chords, melodies, and drum patterns. Training pipelines are included for EnCodec, MusicGen, Multi Band Diffusion, and JASCO, so researchers can train or fine-tune their own variants rather than only run released checkpoints. The library requires Python 3.9 and PyTorch 2.1.0, with ffmpeg for audio handling. The code is MIT licensed, while the released model weights carry a CC-BY-NC 4.0 non-commercial license, a distinction that matters for anyone planning production use. Its main audience is deep learning researchers and developers experimenting with audio generation.