Stability AI and Arm Unveil Stable Audio Open Small: A Lightweight, On-Device Audio Generation Model

Stability AI Introduces Stable Audio Open Small

Stability AI, the AI startup behind the popular image generation model Stable Diffusion, has unveiled its latest innovation in collaboration with chipmaker Arm. The new product, Stable Audio Open Small, is a lightweight audio-generating AI model designed to run efficiently on smartphones and other mobile devices 1

1

.

Technical Specifications and Capabilities

Stable Audio Open Small is a 341 million parameter model optimized for Arm CPUs. It can generate up to 11 seconds of audio in less than 8 seconds, even when running locally on a smartphone 1

1

. The model is particularly adept at creating short audio samples and sound effects, such as drum loops, instrument riffs, and ambient textures 2

2

.

Architecture and Training

The model is based on a latent diffusion architecture using a transformer. It was trained on a dataset of 486,492 audio recordings, all of which are licensed. For text conditioning, a publicly available pre-trained T5 model was utilized. Stability AI also employed the Adversarial Relativistic-Contrastive (ARC) algorithm in post-training to enhance prompt adherence and increase inference speed 2

2

.

Unique Selling Points

What sets Stable Audio Open Small apart is its ability to run offline, unlike many other AI-powered audio generation apps that rely on cloud processing. This feature allows for use in scenarios where internet connectivity is unavailable or real-time generation and responsiveness are crucial 1

1

2

2

.

Ethical Considerations and Limitations

Stability AI claims that Stable Audio Open Small's training set consists entirely of songs from royalty-free audio libraries, specifically the Free Music Archive and Freesound. This approach potentially mitigates intellectual property risks associated with using copyrighted content in training data 1

1

.

However, the model does have limitations. It only supports prompts in English and cannot generate realistic vocals or high-quality songs. Additionally, its performance varies across musical styles, likely due to Western-biased training data 1

1

.

Availability and Licensing

The model weights are available for download on Stability AI's Hugging Face listing, with the code base accessible on GitHub. It's released under the Stability AI Community License, allowing both commercial and non-commercial use. However, there are some restrictions: while free for researchers, hobbyists, and businesses with less than $1 million in annual revenue, larger organizations need to purchase an enterprise license 1

1

2

2

.

Industry Context and Company Background

This release comes at a crucial time for Stability AI. The company recently faced challenges, including financial difficulties and leadership changes. However, it has since appointed a new CEO, added "Titanic" director James Cameron to its board, and released several new image generation models 1

1

.

The collaboration with Arm, announced at Mobile World Congress 2025, represents a strategic move into the mobile AI space, potentially opening new avenues for on-device AI applications 2

2

.