AudioCraft

AudioCraft is Meta AI's open-source framework for generating audio and music, released in August 2023. It bundles three models: MusicGen, which writes music from text descriptions; AudioGen, which creates sound effects from text prompts; and EnCodec, a neural audio compressor that powers both. The code is MIT-licensed on GitHub, while the pretrained weights are for research and non-commercial use. Developers can run it locally or load MusicGen through Hugging Face Transformers. MusicGen was trained on 20,000 hours of music Meta owns or licensed, avoiding scraped data.

Key Features

MusicGen Text-to-Music

MusicGen turns descriptive text prompts into instrumental music, like '80s electric guitar with drum beats. A melody-conditioned version can restyle a melody you supply.

AudioGen Sound Effects

AudioGen creates sound effects from text prompts, like a dog barking or footsteps on wood. Game developers use it to fill virtual worlds cheaply.

EnCodec Compression

EnCodec is a neural codec that compresses audio into tokens a language model can read. Its improved decoder produces higher-quality music with fewer artifacts.

Open Weights and Code

The code is on GitHub under the MIT license, and Meta shares pretrained weights for research. Researchers can train their own models on new datasets.

Hugging Face Integration

MusicGen ships in Hugging Face Transformers and loads in a few lines of Python. Small, medium, and large sizes trade speed for quality.

Frequently Asked Questions (FAQ)

What is the difference between MusicGen and AudioGen?

MusicGen writes music from text prompts, while AudioGen creates sound effects and environmental sounds from text. Both live in the same AudioCraft codebase and share the EnCodec tokenizer.

Can I use AudioCraft music in commercial projects?

Be careful here. The code is MIT-licensed, but the pretrained model weights carry a non-commercial license, so commercial use of generated output needs a different model or a license check.

Do I need a powerful computer to run it?

MusicGen comes in small, medium, and large sizes. The small model runs on a single modest GPU, while the large model needs much more memory for the best quality.

Why did Meta open-source AudioCraft?

Meta says opening the code lets researchers test ways to reduce bias and misuse, since the field needs shared work on responsible audio generation.

Conclusion

AudioCraft is the serious starting point for AI audio research and prototyping. MusicGen and AudioGen cover music and effects, EnCodec handles compression, and the open code invites real experimentation. Just respect the non-commercial weight license. For developers and researchers, it remains the most capable open audio toolkit. It will not replace a composer, but it prototypes ideas fast.