Stable Audio 3 (arxiv.org)

Stable Audio 3 has been announced as a cutting-edge family of fast latent diffusion models designed for variable-length audio generation and editing. These models can create extensive audio tracks in a matter of seconds and come equipped with a novel semantic-acoustic autoencoder that compresses audio into a latent space, enhancing both efficiency and sound fidelity. The included inpainting feature allows for targeted edits and seamless continuation of short audio segments, making it particularly useful for content creators looking to refine specific sounds without generating cumbersome full-length audio. Significantly, Stable Audio 3 democratizes access to high-quality audio generation by providing small and medium model weights that can run on standard consumer hardware. Moreover, thanks to adversarial post-training, these models not only accelerate inference times but also enhance generation quality with fewer steps while maintaining adherence to prompts. With training conducted on licensed and Creative Commons data, this innovation opens new avenues for music and sound creation, positioning it as a valuable tool for the AI/ML community and beyond.