AudioCraft
Meta FAIR's AudioCraft is a PyTorch research library spanning melody- or text-conditioned MusicGen, text-to-sound AudioGen, EnCodec, MAGNeT, JASCO, AudioSeal, and supporting training and inference components.
Why use it
It is selected because one codebase covers music, non-musical effects, neural audio coding, and watermarking: technical teams can compare conditioning methods, measure latency and VRAM, study melodic constraints, and batch candidates into game prototypes, while gaining a more transparent baseline for self-hosted training and inference cost than a single browser demo provides.
Where it fits
Use art, animation, audio, levels, and narrative to make the game easy to read and its world convincing.
What to check
The code is MIT, but official model weights use CC BY-NC 4.0 and cannot be shipped in commercial games; free downloads and published training code do not make generated assets commercially licensed. Public weights remain research releases requiring local GPUs, and downstream use needs review of training data, similarity, bias, watermarking, output quality, and maintenance status.