Stable Audio Open

Stable Audio Open is Stability AI's open counterpart to its commercial Stable Audio service. The 1.0 model (1.21B parameters) combines an audio autoencoder, a T5 text encoder, and a transformer diffusion model operating in latent space to generate variable-length stereo audio up to 47 seconds at 44.1 kHz from a text prompt. It was trained on roughly 486,000 Creative Commons recordings from Freesound and the Free Music Archive, which sidesteps the copyright questions hanging over models trained on commercial music. The tradeoff is scope: it handles drum beats, riffs, ambient beds, and production elements well but is not built for full songs, melodies, or vocals, and the weights carry a non-commercial license. Stability later released Stable Audio 3.0 (May 2026) with open-weight small/medium variants for longer tracks.

Team-reported: competitive FD_openl3 realism scores against state-of-the-art text-to-audio models (paper)