Stable Audio Tools
An open-source Python toolkit generates audio locally and supports model training and fine-tuning, with a Gradio interface and an MIT license.
Stable Audio Tools is a Python toolkit for developers and audio researchers who want to generate audio or train models on their own recordings. It supports work with pretrained models as well as training from scratch.
- Audio generation: Use Stable Audio Open or load compatible models from Hugging Face and local checkpoints for inference on your own hardware.
- Browser interface: Try trained models through a basic Gradio interface, with an optional shareable link for demos.
- Training and fine-tuning: Train a model from scratch, continue an existing training run or adapt pretrained weights to an audio dataset.
- Larger training workloads: Spread training across multiple GPUs or machines, and use gradient accumulation to increase the effective batch size on smaller GPUs.
The toolkit accepts datasets from local audio folders or WebDataset datasets stored in Amazon S3. Its training workflow requires a Weights & Biases account and uses that external service to log outputs and demos.
Stable Audio Tools is open source under MIT. Model licenses remain separate: access to Stable Audio Open through Hugging Face requires accepting that model's terms.
Logotyper
