Stability AI's latest release, Stable Audio 3.0, marks a major advancement in AI-driven music generation, allowing users to create tracks up to six minutes long. This version features a new architecture with a semantic-acoustic autoencoder, improving the flexibility and length of audio outputs compared to earlier models.
The model suite includes four variants, each designed for different music production needs. The two smaller models, Stable Audio 3.0 Small SFX and Stable Audio 3.0 Small, each contain 459 million parameters and can generate tracks up to two minutes long in just 0.44 seconds on an H200 GPU. The Small SFX model focuses on sound effects, while the Small model is intended for shorter music pieces. The Medium variant, with 1.4 billion parameters, allows for audio generation of up to 6 minutes and 20 seconds, completing the process in 1.31 seconds.
The most advanced model, Stable Audio 3.0 Large, features 2.7 billion parameters but is not available with open weights. Access is limited to Stability AI's API, partner fal.ai, or through enterprise licensing on private infrastructures. This model is recognized for its superior musicality, making it suitable for platforms that require high-volume audio generation.
A notable feature of Stable Audio 3.0 is its on-device composition capability. The Small model can create full musical pieces offline without the restrictions of short sample lengths; previous versions were limited to much shorter durations. Stability AI highlights this model's potential for complete creative freedom, enabling users to compose music directly on their devices.
Moreover, Stability AI is offering LoRA training documentation alongside the Small and Medium models. This allows users to fine-tune the models with their own audio libraries, personalizing their music generation experience. Enterprise clients will also receive guided fine-tuning support, helping businesses tailor the technology to their specific requirements.
The models include inpainting features that let users edit segments of a track, modify multiple parts at once, or extend tracks beyond their original length. This flexibility encourages creativity and audio experimentation.
Stability AI's licensing approach also distinguishes it from competitors. Under the Stability AI Community License, users retain ownership of the generated audio and can use it commercially, free of charge up to $1 million in revenue. Businesses that exceed this threshold must seek enterprise licensing for additional coverage and legal protections. The company claims that its licensing structure, supported by partnerships with major music entities like Universal Music Group and Warner Music Group, reduces risks associated with unlicensed data common in rival offerings.
As the audio generation field evolves, Stability AI's Stable Audio 3.0 emerges as a powerful tool for both amateur and professional creators. By emphasizing user control and flexibility, the new models enhance the creative process and help users handle the complexities of commercial audio production confidently. This release may significantly influence the future of audio technology, marking an exciting development for the industry.
