Stability AI released today Stable Audio Open 1.0, a new generative AI model for its audio domain.Stability AI is known for its steady proliferation of text-to-image generative AI technology, but that's just part of the company's portfolio. The company first launched Stable Audio, a text-to-audio generative AI tool, in 2023. The recently released Stable Audio 2.0 improves the clarity and length of generated audio.
Unlike the full version of Stable Audio, which can be used for general commercial purposes and generates audio up to 3 minutes long, Stable Audio Open is more limited in its application scenarios; the goal of Stable Audio Open is to generate short sound clips, not full songs.
As its name suggests, Stable Audio Open is an open model, although it is not open source.Stable Audio Open is made available to users under Stability AI's Non-Commercial Research Community Agreement license, which permits open access to the model but places restrictions on the operations that can be performed using the model.
Zach Evans, Head of Audio Research at Stability AI, said, "Our goal with the launch of Stable Audio Open is to give audio researchers and producers hands-on experience with one of our generative audio models to accelerate the research, adoption, and actual creative use of these incredible new tools. "
What is Stable Audio Open?
Stable Audio Open is a model specifically designed for music production and sound design that optimizes the generation of audio samples such as drum beats, instrumental phrases, ambient sounds, and more. Compared to the commercial version of Stable Audio, Stable Audio Open generates higher quality audio at 47 seconds in length.
Stability AI took a responsible approach to model training, using audio data from FreeSound and free music archives for training to ensure no copyrighted or proprietary material was used.
Stable Audio Open can be fine-tuned by the user
Another major advantage of Stable Audio Open is the ability for users to fine-tune the model to their own custom audio data. For example, a drummer can fine-tune a model based on his or her own drum recording samples to generate new, unique beats.
Stable Audio Open's fine-tuning is implemented through the Stable Audio Tool Library, which is licensed under the actual open source license.Stable Audio Open's model weights are now available on Hugging Face.
Evans said, "The audio research team has been working hard to improve the quality and controllability of the generated audio models. We look forward to further releases of commercial and open models to reflect the progress of our research."