Stable Audio
Stable Audio 3.0 is a family of generative audio models trained on fully licensed data that can generate complete tracks up to six minutes long, with multiple access options including enterprise deployment, API, open weights, and an app.
About Stable Audio
Stable Audio 3.0 is a family of generative audio models that, according to official statements, was trained on fully licensed data and is intended to serve as a foundation for future building by the audio community. Users can generate complete tracks with complex, dynamic musical structures up to six minutes long, and the official team claims industry-leading inference speed with generation completed in seconds. The product is available in multiple ways: an enterprise license allows deployment on your own infrastructure with customization and support options; the Stability AI API provides hosted access to the latest models for integration into apps; Stable Audio 3.0 Small and 3.0 Medium offer open-weight downloads; and there is also the Stable Audio app. The official website does not disclose specific pricing, so costs must be confirmed by contacting the vendor.
Best for
Best use cases
Generating complete background music for videos, games, or other content
Enterprises that need to deploy generative audio models on compute they control
Audio practitioners who want to develop or do research based on open weights
App developers who want to quickly integrate audio generation via an API
Features
Full tracks up to six minutes long
Generate complete tracks with complex, dynamic musical structures, up to six minutes long.
Fast generation
Officially reported to have leading inference speed, completing generation within seconds.
Enterprise-licensed deployment
Deploy on your own infrastructure, with customization and support options available.
Hosted API access
Access the latest models in a hosted way through the Stability AI API, making it easy to integrate into apps.
Open weights
Stable Audio 3.0 Small and 3.0 Medium offer open-weight downloads.
Pros
- Officially stated that the model was trained on fully licensed data
- Supports multiple deployment options, including enterprise self-hosting, API, open weights, and an app
- Can generate complete tracks up to six minutes long with complex, dynamic structures
- Officially stated generation speed is on the order of seconds