Install AudioCraft Locally: A 2023 MusicGen and AudioGen Walkthrough

AudioCraft is Meta’s project for generating music and sound effects. I will show the local workflow on a Mac using MusicGen and AudioGen. This tutorial was recorded in August 2023, so commands and dependencies should be checked against the current AudioCraft repository before following them.

First I create and activate a Conda environment with `conda create -n ac python=3.9` and `conda activate ac`. The environment name appears in the terminal once activation succeeds. I am using macOS; I expect a similar workflow on Windows, but the video does not demonstrate or verify Windows installation.

AudioCraft also needs FFmpeg. On my Mac I use `brew install ffmpeg`; Homebrew reports that it is already installed. The current project documentation also describes installing FFmpeg through Conda, alongside specific Python and PyTorch dependencies.

I copy the repository address from GitHub, run `git clone https://github.com/facebookresearch/audiocraft.git`, and enter it with `cd audiocraft`. In the 2023 walkthrough I install its requirements with `python -m pip install -r requirements.txt`. Installation methods and dependency versions can change, so consult the project README for the current procedure.

To start the MusicGen Gradio demo I run `python -m demos.musicgen_app --share` and open the local URL printed in the terminal. The `--share` option also requests a shareable Gradio link; it is not necessary to assume that every access route is local or private.

The MusicGen screen provides a text box for a musical description and an optional audio input for conditioning on a melody. I enter “90s hip hop West Coast” and submit it. On the first run, model weights must be downloaded, so startup takes longer than on my machine, where the download is already complete.

Generation is taking a while; I believe my run is using the CPU. The interface lets me choose the melody model or small, medium and large variants, adjust duration, and experiment with top-k, top-p and temperature. The optional audio input conditions a melody-capable model; it is not a general promise that any model performs arbitrary audio-to-audio transformation.

I keep the example at ten seconds and listen to the result. I suggest experimenting with the controls and describe the demonstrated interface as allowing a duration up to two minutes. Available limits and extended-generation behavior depend on the version and model; the original MusicGen code uses a 30-second model window with an extension mechanism for longer outputs.

For sound effects I turn to AudioGen. I had trouble using its code and Jupyter notebook on my Mac, so I use a community Gradio interface proposed by cocktailpeanut in AudioCraft pull request 185. My own demonstration worked, but that does not guarantee the same setup will work on every computer.

At recording time the pull request had not been merged, so I fetch it with `git fetch origin pull/185/head:PR185`, then switch with `git checkout PR185`. I expected it might be merged later, but the project page still lists this proposal as open; it should not be assumed to be part of the main branch.

From that branch I run `python -m demos.audiogen_app --share`, open the local URL, and enter “dog barking.” The first use needs to download the AudioGen model; mine was already cached. The resulting sound resembles a small dog, and I find it convincing.

The demonstration is free to try locally, but licensing deserves a distinction: AudioCraft repository code is MIT-licensed, while the released model weights are under CC BY-NC 4.0. Check the applicable terms before using the weights, especially for commercial purposes. I encourage viewers to experiment with the audio settings and results.