AudioCraft

AudioCraft is Meta AI's open research codebase for generative audio, unifying three models — MusicGen for text-to-music generation, AudioGen for text-to-sound-effect generation, and EnCodec, the neural audio codec that underlies both. Rather than the more complex pipelines used in prior generative-audio work, AudioCraft represents audio as streams of discrete tokens produced by EnCodec and models them with a single autoregressive language model, using a token-interleaving pattern to capture long-term dependencies while keeping the architecture simple. MusicGen turns text prompts into diverse, longer-form music samples, while AudioGen focuses on generating environmental and sound-effect audio from text descriptions. The code, model cards, and technical writeups are published openly for researchers and developers to build on.