AudioCraft is a PyTorch-based library developed by Facebook Research that pushes the boundaries of what’s possible in AI-powered audio generation. The library contains inference and training code for two cutting-edge models – AudioGen for text-to-sound generation and MusicGen for text-to-music generation. MusicGen, in particular, can create coherent and high-fidelity music pieces from detailed text descriptions, allowing users to generate custom audio tracks with remarkable ease. The library also includes EnCodec, a neural audio codec that can compress audio files to a fraction of their original size without significant quality loss. AudioCraft provides researchers and developers with the tools to explore the frontiers of generative audio AI.