AudioCraft

AudioCraft is an open-source PyTorch library developed by Meta AI for generative audio applications, encompassing state-of-the-art models for music generation, sound effects, and neural audio compression. The framework unifies state-of-the-art models like MusicGen and AudioGen into a standardized architecture powered by an autoregressive language model and the EnCodec neural audio compressor. EnCodec maps raw audio waveforms into discrete tokens, allowing the language model to process parallel streams of audio representation efficiently before decoding them back into high-fidelity sound. Content creators, researchers, and developers can produce realistic sound effects or complex musical arrangements using descriptive text prompts. AudioCraft offers modular components that allow researchers to train custom generative models on proprietary datasets or fine-tune existing checkpoints for specific sonic domains. Despite its flexibility, running training or high-resolution audio inference requires substantial GPU computational resources, and fine-grained control over specific musical parameters like tempo or key changes remains limited. Nonetheless, AudioCraft provides a robust foundation for building advanced generative audio pipelines and experimental sound synthesis workflows.