facebookresearch/audiocraft is a PyTorch library for deep learning research on audio processing and generation. The repository description and README state that the project includes inference and training code for AudioGen and MusicGen. It is distributed under the MIT license and has received 23,751 stars on GitHub. The repository lists Jupyter Notebook as its primary language, and its creation date is June 8, 2023. No public official release version could be confirmed.
To start with what it does, this library covers generative models for ambient sound and music, and provides PyTorch components for audio deep learning research along with model training pipelines. At its center are EnCodec and MusicGen. EnCodec is an audio compressor and tokenizer that turns waveforms into symbol units that models can handle easily. MusicGen is a language model (LM) that generates music conditioned on text and melody, and the repository description introduces it as a simple and controllable music generation model. The README states that, in addition to EnCodec, it covers Multi Band Diffusion, MAGNeT, AudioSeal, MusicGen Style, and JASCO.
There are caveats worth noting. With no published release information, it is difficult to track what state the AudioGen and MusicGen code is in, and because GitHub shows the primary language as Jupyter Notebook, this information alone does not determine whether the core of the code is a package or examples. The 23,751 stars are only a measure of interest and do not guarantee quality. There are no performance comparison figures, and among the scored items only momentum is given, at a low 3.1 out of 10, while the freshness, technical, and quality scores are not given. Judging only by the documented scope, it is close to a bundle of research code worth examining together with EnCodec and MusicGen when starting audio generation research.