Stability AI Releases Stable Audio 3: A Family of Fast Latent Diffusion Models for Audio Generation and Editing

Stability AI has unveiled Stable Audio 3, a new family of latent diffusion models. These models are specifically designed for generating instrumental music and sound effects. The release emphasizes accessibility with open-weight versions for small and medium variants.These models are optimized for performance across different hardware. The small variant operates efficiently on a MacBook Pro M4 CPU, while the medium variant is compatible with consumer GPUs equipped with 8 GB of VRAM.Stable Audio 3 utilizes a sophisticated three-stage training pipeline, which includes flow matching, distillation warmup, and adversarial post-training. This methodology enables the generation of high-quality stereo audio at 44.1 kHz.The efficacy of Stable Audio 3 is highlighted by its performance on benchmarks. The medium variant achieved an FAD score of 0.369 on the BBC Sound Effects benchmark (at 5 seconds), outperforming all other open-weight baselines evaluated in the accompanying paper.