Meta open-sources AudioCraft for AI music and sound effects

AudioCraft is Meta's open-source powerhouse that turns simple text prompts into realistic music, sound effects, and audio compression models in one place.

While generative AI has exploded across text and images, audio generation has lagged behind. Modeling sound requires capturing complex signals at multiple scales, from fine-grained acoustic waves to long-range melodic structures across different instruments. Existing systems have been complicated and largely locked down behind proprietary walls.

AudioCraft changes that by packaging Meta's long-running audio research into a single, modular codebase with fully open weights.

What it does

AudioCraft brings together three distinct models designed to work in tandem.

MusicGen handles music generation. Trained on Meta-owned and specifically licensed music catalogs, it transforms descriptive prompts about genre, mood, or instrumentation into consistent musical tracks.

AudioGen focuses on environmental and Foley sounds. Meta released pre-trained weights for AudioGen that were trained on public sound effects, allowing users to synthesize realistic ambient sounds like dogs barking, cars honking, or footsteps walking across a wooden floor.

EnCodec is the neural audio codec tying everything together. The release includes an upgraded EnCodec decoder that produces higher-fidelity audio with fewer artifacts, making generated samples sound noticeably crisper and less synthetic.

Under the hood

Instead of treating audio generation as disconnected problems, AudioCraft provides a unified foundation for sound design, compression, and generation.

Meta released all model weights and code, giving developers the entire recipe rather than just an inference demo. Researchers can now take their own private datasets and train or fine-tune custom audio models directly on top of Meta's foundation.

Whether you are building a specialized Foley engine for video editing or pushing the boundaries of neural audio compression, everything lives inside the same repository.

Why it matters

Meta is replicating its open-source playbook in a domain where proprietary platforms have kept their models tightly under wraps.

For indie game devs, video creators, and musicians, AudioCraft provides an accessible brainstorming tool to rapidly draft soundscapes and demo tracks without touching an instrument. Meta compares MusicGen's trajectory to early synthesizers, predicting it could evolve into an entirely new instrument once granular controls mature.

We think open-sourcing the entire stack is a brilliant move that will crowd-source the hardest engineering problems in AI audio. While prompt-to-audio still lacks the surgical precision professional sound designers need for final production, AudioCraft dramatically lowers the barrier to entry for everyday prototyping.