Stable Audio Introduction

Stable Audio employs a sophisticated latent diffusion model architecture consisting of a variational autoencoder (VAE), a text encoder, and a U-Net-based conditioned diffusion model. For text-to-audio generation, users input descriptive prompts which the model interprets to create corresponding audio. The audio-to-audio feature allows users to upload existing audio samples and transform them using natural language prompts, enabling style transfers and variations. The platform also incorporates innovative conditioning on text metadata, audio duration, and start time, allowing for precise control over the generated audio's content and length. This approach overcomes limitations of traditional fixed-size audio outputs, enabling the creation of coherent musical structures with intros, developments, and outros.