Stable Audio Open

Behind Stability AI's audio models sits stable-audio-tools, the company's open-source Python codebase for training and running generative audio systems. It covers several model families: neural audio autoencoders, latent diffusion models for unconditional, conditional, and inpainting generation, diffusion autoencoders, and audio language models. Training is built on PyTorch Lightning with multi-GPU and multi-node support, DeepSpeed ZeRO stage 2, and Weights & Biases logging, driven by JSON configuration files that define model hyperparameters and datasets, which can be local audio folders or WebDataset shards on S3. Trained checkpoints include optimizer and discriminator state, and a provided unwrapping script extracts clean inference models for deployment or fine-tuning from pretrained weights such as Stable Audio Open. A Gradio interface makes it easy to test models interactively. The toolkit requires Python 3.10 and PyTorch 2.5 or newer for Flash Attention support. Its users are mostly researchers and audio ML engineers fine-tuning or training text-to-audio models rather than end users looking for a finished application.