Stable Audio Open: audio generator — what it does and what you need to run it

Stable Audio Open is an open source audio generator from Stability AI. It solves the text-to-audio task: it turns a text description into an audio file. The tool suits those who need audio generation from text and are willing to work with the code and model weights directly.

What it does

The authors describe the project as generative models for conditional audio generation. The task the model solves is text-to-audio: text goes in, audio comes out. No other modes, such as speech generation or music from notes, are stated in the fact sheet.

The code is written in Python. The latest change to the repository was on August 9, 2026, and the repository itself was created on May 23, 2023.

What you need to run it

You need the stable-audio-tools library to run it. The largest model weights file is 4.5 GB, and all weights files together are 14.6 GB. This is worth considering when planning disk space.

As described by the author, development is done in Python 3.10. The repository description states: "Requires PyTorch 2.5 or later for Flash Attention and Flex Attention support. Development for the repo is done in Python 3.10." That is, PyTorch 2.5 or newer is required for Flash Attention and Flex Attention support.

The code is distributed under the MIT license. The weights license is listed as other — the authors do not disclose specific terms in the fact sheet.

Who it suits

The tool is aimed at those who know how to work with Python and are not afraid to run models through code. The open source and available weights allow you to understand how the model works and, if needed, refine it for your own tasks.

For those looking for a ready-made application with an interface, this option may not fit: the fact sheet has no information about a graphical interface or web service. Also, consider the size of the weights — 14.6 GB in total, which can be significant for weaker machines.

Stable Audio Open is a tool for developers and researchers who need text-to-audio generation and are willing to work with the code and weights directly. The code is open, the model is available, but running it requires technical skills and free disk space.