Stable Audio Tools is Stability AI's open-source codebase for training and running generative audio models, and it is the reference implementation behind the company's Stable Audio Open release. Distributed under the permissive MIT license, the repository packages the full stack — model definitions, training loops, inference scripts, and a ready-to-use interface — into a single toolkit so that researchers and developers can generate, fine-tune, and study text-conditioned audio models on their own hardware. With more than 3,700 GitHub stars, it has become a common starting point for anyone building open audio generation systems. ## A Full-Stack Toolkit for Generative Audio Rather than shipping a single model, Stable Audio Tools provides the machinery to work with an entire family of audio generation architectures. The codebase covers conditional audio generation from text prompts, autoencoder training, and latent diffusion, all driven by configuration files. A model config describes the architecture while a dataset config describes the training data, which keeps experiments reproducible and makes it straightforward to swap components. Dependency management is handled through uv for fast, reproducible installs, and Flash Attention is supported for higher throughput on capable GPUs. ## Built Around Latent Diffusion The toolkit is organized around a latent diffusion approach: an autoencoder compresses raw audio into a latent space, and a diffusion model is trained to generate within that space. This pretransform mechanism lets a trained autoencoder be reused as the front end for a separate latent diffusion model, and the repository includes utilities for unwrapping checkpoints so that an autoencoder, decoder, or diffusion model can be tested and recombined independently. The design mirrors how modern image diffusion systems are structured, applied here to waveform and spectral audio data. ## Inference, Training, and Fine-Tuning A bundled Gradio interface makes it easy to try a trained model interactively — running run_gradio.py with a Hugging Face model name such as stabilityai/stable-audio-open-1.0 launches a local web demo with options for shareable links and basic authentication. On the training side, the project uses PyTorch Lightning to support multi-GPU and multi-node runs, with Weights & Biases integration for logging outputs and audio demos. Checkpoints created during training embed a training wrapper holding optimizer states and EMA copies, and a provided unwrap step strips that wrapper down to a lean inference checkpoint. Fine-tuning is supported by continuing a run from a pre-trained checkpoint, including partial initialization for modified configurations. ## Open Weights and the Stable Audio Open Model The most visible payoff of the toolkit is Stable Audio Open, the openly licensed model whose weights are hosted on Hugging Face and loaded directly through this codebase. The released model generates variable-length 44.1 kHz stereo audio from natural-language descriptions, making it suited to sound effects, instrument and drum samples, ambient textures, and short musical ideas. Because both the model and the surrounding training code are open, teams can fine-tune on their own licensed datasets rather than relying on a closed API, which is a meaningful difference for production and research use. ## Considerations Stable Audio Tools is firmly a developer and researcher tool rather than a polished consumer application: it expects familiarity with Python, PyTorch, configuration files, and GPU setup, and serious training runs require substantial compute plus a Weights & Biases account. The Gradio demo lowers the barrier for inference, but newcomers without machine-learning experience will find the learning curve steep. Users must also accept the model terms on Hugging Face before downloading weights and remain mindful of dataset licensing when fine-tuning. For those building open, self-hosted audio generation pipelines, however, the combination of open code and open weights under MIT makes it one of the most flexible options available.