Audiocraft

Open-source library behind MusicGen, AudioGen and the EnCodec codec

AI Audio & MusicFree planOverseasβ˜…β˜…β˜…β˜†β˜† 3.0

What is Audiocraft?

Audiocraft is Meta's research library for audio generation and the home of MusicGen, AudioGen and the EnCodec neural audio codec. It bundles training and inference code, so you can generate music from text or a melody, produce sound effects and environmental audio, and compress or reconstruct audio at low bit rates. Running it locally needs a GPU and a Python environment, but it gives full control: fine-tune on your own material, batch-generate, and read the pipeline end to end. The code is permissively licensed, while the pretrained weights carry separate non-commercial terms that matter if you plan to ship something.

Last updated: 2026-09-20. This site only provides an index; for exact features, pricing, and licensing, see the official website.

Key features

  • MusicGen text-to-music generation with optional melody conditioning
  • AudioGen text-to-sound-effects and environmental audio generation
  • EnCodec neural audio codec for high-quality low-bitrate compression
  • Training and fine-tuning scripts shipped with the inference code
  • Pretrained checkpoints in several sizes for different hardware
  • Python API and command-line tools for batch generation

Pros & cons

Strengths

  • One library covers music, sound effects and audio codecs
  • The code is permissively licensed and freely modifiable
  • Strong documentation and an active research community around it

Watch out for

  • Pretrained weights are non-commercial, which limits product use
  • Needs a GPU and a working Python environment to be practical
  • It is a raw library, with no interface for non-developers

Best for & use cases

audio research, model fine-tuning, game sound prototyping and ml projects

If you're comparing similar products, check the alternatives below, or browse all tools in the AI Audio & Music category.

FAQ

Is the licence really non-commercial?

The code is permissively licensed, but the pretrained model weights are released under research terms that restrict commercial use. Treat the two separately and read the model card for the checkpoint you intend to use.

Which models are included?

MusicGen for music, AudioGen for sound effects and environmental audio, EnCodec for compression and reconstruction, plus the training code that ties them together. Each has checkpoints in several sizes.

Can I fine-tune it on my own audio?

Yes, the repository includes training and fine-tuning scripts, which is a common reason teams pick it over a hosted service. Expect to supply a suitable dataset and to manage GPU memory carefully.