audiocraft
facebookresearch/audiocraft · Jupyter Notebook · MIT
MusicGen and companion models for text-to-music
AudioCraft is Meta’s research toolkit behind MusicGen. It is still one of the easiest ways to try controllable text-to-music locally with solid docs.
Score
Setup at a glance
pip install -U audiocraft Deep dive
AudioCraft is Meta's codebase for neural audio generation: it hosts the family of models that defined the text-to-audio wave — MusicGen for text-to-music, AudioGen for text-to-sound, EnCodec for high-fidelity neural compression, and Multi Band Diffusion as an EnCodec-compatible decoder.
Each model is documented in its own guide with training and inference code, which is what made AudioCraft the reference implementation a generation of audio projects forked from.
It is a research codebase: expect to configure checkpoints and environments yourself, and to read the per-model docs before running anything serious.