Fast Timing-Conditioned Latent Audio Diffusion

TL;DR

This paper introduces Stable Audio, a fast and efficient latent diffusion model capable of generating long-form, variable-length stereo audio at 44.1kHz from text prompts, with fine control over content and duration.

Contribution

It presents a novel approach using timing-conditioned latent diffusion for high-quality, long-form audio generation that outperforms existing models in speed and structure.

Findings

Capable of generating 95-second stereo audio in 8 seconds.

Achieves top performance on public text-to-music and -audio benchmarks.

Can produce structured music with stereo effects, unlike previous models.

Abstract

Generating long-form 44.1kHz stereo audio from text prompts can be computationally demanding. Further, most previous works do not tackle that music and sound effects naturally vary in their duration. Our research focuses on the efficient generation of long-form, variable-length stereo music and sounds at 44.1kHz using text prompts with a generative model. Stable Audio is based on latent diffusion, with its latent defined by a fully-convolutional variational autoencoder. It is conditioned on text prompts as well as timing embeddings, allowing for fine control over both the content and length of the generated music and sounds. Stable Audio is capable of rendering stereo signals of up to 95 sec at 44.1kHz in 8 sec on an A100 GPU. Despite its compute efficiency and fast inference, it is one of the best in two public text-to-music and -audio benchmarks and, differently from state-of-the-art models, can generate music with structure and stereo sounds.