RepoPicks Radar

Updated every day around 23:30 UTC

Guides

AudioCraft

facebookresearch/audiocraftReviewed by AI

Checked by RepoPicks Radar on

What is AudioCraft?

AudioCraft is the research code behind Meta's audio models: MusicGen (music from a text description, and from a melody you hum), AudioGen (sound effects from text), EnCodec (audio compression), MAGNeT, JASCO and AudioSeal watermarking. It is a Python library with example notebooks, not a program with a window: you install it with pip and write or run code. It expects Python 3.9, PyTorch 2.1.0, ffmpeg and an NVIDIA graphics card. The code is MIT, but the trained models are licensed for non-commercial use only.

AudioCraft is free and open source (MIT) and can replace Suno, Udio, and Epidemic Sound.

How easy is it to install AudioCraft?

E · developers only · Windows

  • Windows E
  • Linux E

There is nothing to open on either Windows or Linux: the README only shows pip commands for a Python library, plus installing ffmpeg, and the examples are notebooks and training recipes. You also have to get a specific old combination to work — Python 3.9 with PyTorch 2.1.0 on a CUDA graphics card — which is the hard part on a recent machine.

Coming soon: install in 1 click with the RepoPicks app.

AudioCraft on GitHub

Stars
23,839on Nov 10, 2026
Downloads
—all releases
Growth
+19stars in 6 days

Stars over time

23,82023,82523,83023,83523,840 Nov 4Nov 6Nov 8Nov 10 23,82023,82523,83023,83523,840 Nov 4Nov 7Nov 10

Facts

Runs on
Windows, Linux
You need to install
Python, NVIDIA graphics card, FFmpeg
To build from source
Python, NVIDIA graphics card, FFmpeg
Disk space
1.0 GB–4.6 GB · estimate
Accounts or keys
None
Cost
Free · No charge. The code is MIT, but the trained models are CC-BY-NC 4.0: non-commercial use only, so what you generate cannot be sold or used in paid work.
License
MIT (Permissive)
Last change on GitHub
Mar 13, 2025

AudioCraft risk signals (6)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-5pcm-hx3q-hm94

    automatic check

  • Attentionmodels are non commercial only

    README License section: "The models weights in this repository are released under the CC-BY-NC 4.0 license as found in the LICENSE_weights file." The MIT licence covers only the code, so music or sound made with the released models may not be used commercially.

    AI review

  • Attentionno changes for over a year

    Radar's check records the last commit on 2025-03-13, 600 days before the check, with no GitHub releases ever and 401 open issues. Nobody is fixing install problems or supporting newer graphics cards.

    AI review

  • Attentionpinned to old versions

    README: "AudioCraft requires Python 3.9, PyTorch 2.1.0"; requirements.txt pins "torch==2.1.0", "torchvision==0.16.0", "torchtext==0.16.0", "numpy<2.0.0", "xformers<0.0.23", "spacy==3.7.6". Those versions are from 2023, which is also why Radar's check found known vulnerabilities in the pinned torch, and recent NVIDIA cards need a newer PyTorch than this allows.

    AI review

  • Infomodels downloaded at first use

    README FAQ "Where are the models stored?": the weights come from Hugging Face into a cache folder ("AUDIOCRAFT_CACHE_DIR"), and musicgen-melody also pulls Demucs through Torch Hub. So the real download and disk use happen the first time you generate something, beyond the 1050 MB to 4750 MB Radar estimates.

    AI review

  • Inforesearch code not an application

    README: "AudioCraft is a PyTorch library for deep learning research on audio generation", and the repository holds training recipes (egs/, config/) and demo notebooks (jasco_demo.ipynb, demos/). Non-programmers are better served by one of the ready apps built on top of MusicGen.

    AI review

Evidence is shown in its original language.

What Radar could not check about AudioCraft

  • Whether the 2023 version pins still install on a current machine and a current graphics card.
  • How much graphics memory each model needs, and how long one minute of music takes.
  • Real disk use after the models are downloaded.
  • What the models were trained on, which matters for anyone worrying about the rights to the music that comes out.
  • Whether the examples run on Windows, where the README gives no specific instructions.

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub