MetaAudioCraft is Meta's open-source framework for generating audio with AI: music, sound effects, and the compression technology underneath both. Released by Meta's AI research team and freely available to anyone, it bundles a set of models that turn a written description into sound: type "an upbeat 8-bit chiptune with a driving bassline" or "rain falling on a tin roof with distant thunder," and AudioCraft produces audio to match. It is less a polished consumer app than a powerful, open toolkit that a wave of music and audio products has been built on top of.
What makes AudioCraft significant is that it is open source. Like Meta's Llama language models, the audio models are published for researchers and developers to download, run, study, and build on for free. That openness turned AudioCraft into foundational infrastructure for AI audio, the engine behind many text-to-music and text-to-sound tools, rather than a single closed product locked behind a subscription.
This guide covers everything that matters about AudioCraft in 2026: the three models inside it, what each one does, how to use the framework, who it is really for, how it compares to consumer music generators, and the limitations to keep in mind. By the end you will know whether AudioCraft is the right foundation for your audio work.
What Is AudioCraft?
AudioCraft is a single, open-source code base for generative audio, built by Meta. Rather than one model, it is a library that brings together three of them (MusicGen for music, AudioGen for sound effects, and EnCodec for high-quality audio compression) so that everything you need to generate, process, and encode audio with AI lives in one place. Developers and researchers download it, run it on their own hardware, and use it as the basis for audio applications.
The core idea is text-conditioned generation: you describe the audio you want in words, and the models produce it. MusicGen can also be guided by a melody, letting you hum or supply a tune that shapes the output. Because the models are open and the code is public, you have full control, to fine-tune, to integrate, and to run privately, that no closed API offers.
AudioCraft is aimed primarily at developers, researchers, and technically inclined creators rather than at someone who just wants to click a button and get a song. It is the engine, not the dashboard, though plenty of friendlier tools and community forks have wrapped it in easier interfaces.
The Three Models Inside AudioCraft
AudioCraft's power comes from the three complementary models it ships.
| Model | What it does |
|---|---|
| MusicGen | Generates music from a text prompt, optionally guided by a reference melody. Trained on Meta-owned and licensed music. |
| AudioGen | Generates sound effects and environmental audio from text: footsteps, rain, traffic, animal sounds, and more. |
| EnCodec | A neural audio compressor and tokenizer that encodes sound efficiently, enabling higher-quality generation with fewer artifacts. |
EnCodec is the quiet hero: it compresses audio into a compact set of tokens that the generative models work with, and an improved decoder is what lets MusicGen and AudioGen produce cleaner results. MusicGen itself comes in several sizes, from a few hundred million parameters up to a few billion, so you can trade quality against the hardware and speed you have available.
How to Use AudioCraft
Because AudioCraft is a code library rather than a hosted service, using it looks a little different from a typical AI app.
- Get the framework: the code and model weights are published openly for download.
- Run it on your hardware: load a model (MusicGen, AudioGen, or both); the larger models benefit from a capable GPU.
- Prompt it: describe the music or sound you want in text, and optionally supply a melody to guide MusicGen.
- Generate and refine: produce audio, adjust your prompt or model size, and iterate toward the result you want.
- Integrate or fine-tune: build it into your own application, or fine-tune the models on your own data.
For those who would rather not set up the code themselves, the models are widely available through community interfaces and hosted demos, and active forks add features like stereo output and longer generation. But the canonical, most flexible way to use AudioCraft is to run the open framework directly.
Who AudioCraft Is For
AudioCraft's open, technical nature makes it a fit for some users and not others.
Developers and Builders
The primary audience is developers building audio features into their own products (music tools, game audio systems, creative apps) who want a free, open foundation they can run, customize, and ship without per-use fees or vendor lock-in.
Researchers
AudioCraft was released explicitly for the research community. Because the models and code are open, researchers can study how generative audio works, benchmark against it, and advance the field, something closed models do not permit.
Technical Creators and Hobbyists
Musicians, game makers, and tinkerers comfortable with code use AudioCraft to generate royalty-free backing tracks, sound effects, and sonic experiments locally and for free, with full control over the process.
How AudioCraft Compares
AudioCraft sits opposite the polished, subscription-based consumer music generators. The trade-off is openness and control versus convenience.
| AudioCraft | Consumer music generators | |
|---|---|---|
| Open source | Yes, free to run, fine-tune, and integrate | No, closed subscription apps |
| Ease of use | Technical, a code library to run yourself | Polished, click-to-generate interfaces |
| Cost | Free (you provide compute) | Typically freemium with paid tiers |
| Control & privacy | Full, runs on your hardware | Limited, runs on the provider's servers |
The short version: choose AudioCraft when you want a free, open, fully controllable foundation and you are comfortable with code, especially for building products or doing research. If you simply want to generate a finished song through a friendly interface with no setup, a dedicated consumer music app will be easier, at the cost of openness and control.
Pricing
AudioCraft is free and open source. There is no subscription and no per-generation fee; you download the models and run them on your own hardware. The only real cost is the compute you supply: the larger, higher-quality models benefit from a capable GPU, while smaller models run on more modest setups. For anyone generating audio at volume, that free, run-it-yourself model is a major advantage over per-use commercial services.
Limitations to Keep in Mind
| Limitation | What to know |
|---|---|
| Technical to use | AudioCraft is a code library, not a polished app. Using it directly requires comfort with running models and managing dependencies. |
| Needs compute | The larger, best-sounding models want a capable GPU; smaller models trade quality for the ability to run on modest hardware. |
| Short generation lengths | The base models generate relatively short clips; longer pieces need community forks or additional engineering. |
| Quality vs. top apps | Output is impressive for an open model but may trail the latest closed, consumer-focused music generators on polish. |
| Licensing care | Check the model and music licensing terms before commercial use to ensure your generated audio is cleared for your purpose. |
Final Verdict
AudioCraft is the open foundation of AI audio generation. By bundling MusicGen, AudioGen, and EnCodec into one free, open-source framework, Meta gave developers and researchers a powerful, fully controllable toolkit for creating music and sound, one that has quietly powered much of the AI audio ecosystem. For building, customizing, and experimenting, its openness is unmatched.
It is not a click-and-go consumer app, and it asks for technical comfort and some compute, but for anyone who wants a free, private, open engine for generative audio, AudioCraft is a standout. It comes from the same open philosophy as Meta's Llama models, and you can browse more free AI tools to round out your creative stack.
Frequently asked questions
Is AudioCraft free?
Yes. AudioCraft is fully open source and free to download and run on your own hardware, with no subscription or per-generation fees. The only cost is the compute you provide, and larger models benefit from a capable GPU.
What can AudioCraft generate?
It generates music from text via its MusicGen model and sound effects and environmental audio via AudioGen, with the EnCodec model providing the high-quality audio compression underneath. MusicGen can also be guided by a reference melody.
Who makes AudioCraft?
AudioCraft is built and open-sourced by Meta's AI research team, the same group behind the open Llama language models. It was released for researchers and developers to use and build on freely.
Do I need to be a developer to use AudioCraft?
To use it directly, mostly yes, since it is a code library rather than a polished app. However, the models are also available through community interfaces and hosted demos that make them easier to try without setting up the code yourself.
What are the three AudioCraft models?
MusicGen generates music from text or melody, AudioGen generates sound effects and environmental audio from text, and EnCodec is a neural audio compressor that encodes sound efficiently to enable higher-quality generation with fewer artifacts.
Can I use AudioCraft music commercially?
AudioCraft is open source, but you should check the specific model and training-data licensing terms before commercial use to ensure your generated audio is cleared for your intended purpose.