Meta releases AudioCraft (MusicGen, AudioGen, EnCodec)

Meta released AudioCraft, a framework bundling three separate generative audio models: MusicGen, which composes music from text descriptions using training data Meta owned or specifically licensed; AudioGen, which generates sound effects and ambient audio — footsteps, a dog barking, cars honking — from text, trained on public sound-effect libraries; and EnCodec, an improved neural audio codec that reduced artefacts in the generated output and underpinned the other two.

Unlike some earlier generative-audio research, which had been published as papers without runnable code, Meta released model weights and code for all three components. The company framed the tools as aids for musicians and sound designers to brainstorm and iterate rather than as replacements for composition, and pointed to the open release as a way to let the wider research and creator community build on the work rather than keeping it inside the company’s own products.

The release extended a pattern already established with Llama’s open weights that same year: Meta positioned itself as the major lab most willing to publish usable model weights across modalities, at a time when OpenAI, Google and Anthropic kept their comparable systems closed. AudioCraft’s components were rapidly adopted into open-source tooling and follow-on research on text-to-audio generation, and EnCodec in particular was reused as a building block in later audio and speech models trained by other groups.