Music generation: Stability AI launches Stable Audio 3.0 model family with open weights

Stability AI has released Stable Audio 3.0, a new family of generative audio models trained on fully licensed music data. The release includes four models designed for different devices and use cases, three of which are available as open-weight models that anyone can download and modify.

The four models are:

• Small SFX: generates sound effects on consumer devices such as phones and laptops
• Small: generates full music compositions on consumer devices, up to two minutes long
• Medium: generates longer tracks of up to six minutes and twenty seconds with greater musical structure
• Large: built for professional platforms requiring fast, high-volume generation

The small and medium models are available for free download on Hugging Face. The large model is only accessible through the Stability AI API or paid self-hosting services. Organizations with more than one million dollars in annual revenue must obtain an enterprise license for commercial use.

According to Stability AI, the new models represent a significant technical step forward. The previous open model, Stable Audio Open, could generate audio of up to 47 seconds. The small model in the new family now generates up to two minutes, while the medium and large models reach more than six minutes. Stability AI claims the small model is the first capable of producing complete music compositions on a consumer device.

The company highlights variable-length audio generation as a key feature. Users can specify the exact duration of a track, down to the second. The models also support audio inpainting, which allows users to modify specific sections of a track or extend a composition beyond its original endpoint.

Stability AI says it is also releasing documentation for LoRA training, a technique that allows users to fine-tune models on their own audio libraries. This method was first popularized in image generation and is now being applied to audio models.

All models in the family were trained on fully licensed data, according to Stability AI. The company has existing partnerships with Universal Music Group and Warner Music Group. Licensing has become a notable issue in the AI music space, as competitors Suno and Udio are currently involved in legal disputes over the use of unlicensed music in training data.

Sources: Stability AI, TechCrunch